Build Your AI Research Department

Build Your AI Research Department: The Research Desk

Stop asking AI random questions. Give it a research assignment.

A one-day class where you build a research desk inside ChatGPT or Claude that takes an assignment, gathers sources into a ledger, survives a citation audit you run by hand, counts how many of its sources are actually independent, maps where they disagree, goes past a table of critics, and hands you a one-page brief with a confidence on every claim. Then it watches the question for you. You build it on your real question, run it twice, then build a second desk for a different question with nobody helping.

No code. The whole desk is a Project (or custom GPT) with instructions, plus the deep-research mode you already pay for. Your job is the part the tool can't do: deciding what the research is for, opening the links, and ruling on what the critics say. You're the authority on your own work. The desk is your staff.

Part of AI at Work. Uses the Spec Card.

Who this is for

• Fundraising and development: prospect research, funder priorities, the foundation that just changed its guidelines.
• Marketing and comms: what a competitor actually launched versus what three blogs say they launched.
• Policy and advocacy: what a rule says, who's suing over it, and what happens next.
• Recruiting: a market, a role, a pay range you can source.
• Compliance: what changed in the regulation, in the regulation's own words.
• Consulting: the background brief you write for every new client.
• Organizers: who decides, what they've said on the record, and who's funding the other side.

If you've ever pasted an AI answer into an email and then found out one of its links didn't say the thing, this is your class. In Anthropic's Economic Index, learning-type interactions rose from about 23% to about 28% of Claude conversations (as of March 2025), and educational instruction and library work was 15% of Claude.ai use (as of January 2026). People already ask AI to find things out. What most people don't have yet is the assigned, audited, re-run version. (The Index has no figures for competitor, regulatory, donor or market research specifically, so we won't pretend it does.)

The machine you leave with

INTAKE → STRUCTURE → CHECK → GATE → ACT → LEDGER → RHYTHM
the        source       citation   you sign   the brief,   ledger +    the watch
assignment ledger,      audit,     the brief  to the       audit +     list re-runs
(names the fields fixed independence           person who   diff of     and reports
decision)  first        count,                 decides      two runs    only what
                        critics                                         changed

The Research Desk: a ChatGPT Project or custom GPT, or a Claude Project, whose instructions hold your spec card, the Research Assignment template, your source rules and three critic personas. You hand it an assignment. It runs deep research and returns a source ledger (claim, source, date, primary or secondhand, who's behind it, the quoted sentence). You audit ten citations by hand. It builds a contradiction table and an uncertainty map, convenes the critics, and drafts a brief in which every claim carries a confidence and a ledger row. You rule on the critics and sign the brief. A scheduled task re-runs the watch list and reports only what changed.

The metaphor: the newsroom

The assignment desk decides what a story is for before anyone leaves the building. Reporters go out and come back with notes. The fact-check desk opens every source. The debate room is where two reporters with conflicting sources fight it out, and where three outlets quoting one press release get counted as one. The critics' table reads the draft as the people it'll make angry. The front page runs only what survived, with what we don't know printed alongside it. The beat is the reporter who keeps watching after the story runs. The atlas is drawn this way: /x/ai-at-work-research-map.

The day (10am–4pm, with lunch)

The six blocks

Standards

Capability ids from _capabilities.yml. Only assessed standards are assessed, and they're assessed by looking at the artifact, never by a quiz.

Portfolio: one link, five boxes

Submit one shared link: the custom GPT, or the shared Claude Project (or an artifact published from it). The facilitator checks five things against that link:

• [ ] Spec card in the instructions, every slot filled, with the Research Assignment and source rules under it
• [ ] Two real runs visible: the first ledger and the second (later, other product, or sibling question), and the diff between them
• [ ] One failure the check caught: a citation audit row marked wrong claim, dead or secondhand, and the source rule it added
• [ ] The human gate named: who signs the brief, and the critic rulings they made
• [ ] A second desk for a different question in a different domain, with one real run, and the watch list scheduled on the first

Prerequisites

• A laptop and a browser.
• A paid ChatGPT or Claude plan is strongly recommended, because deep research / Research mode and Projects are limited or missing on free tiers (as of Nov 2026; facilitator re-checks before class). A free NotebookLM account helps for the audit.
• A real question from your work, and the decision it feeds. Write it on a sticky note before you arrive: "Someone has to decide ___ by ___." If you can't share yours, use the Practice Pack.
• No code. No prior AI class. If you've taken another AI at Work class, Block 1 will go faster.

What we don't do

• We don't let it write the conclusion first. The brief is the last thing written, not the first.
• We don't take "it gave me links" as an audit. You open them.
• We don't treat both sides as equal because there are two. False balance is a failure the critics' table is built to catch, the same as one-sidedness.
• Nothing leaves without a person signing it. The desk drafts.

How it relates to other classes

• Practical Propaganda teaches you to check your own claims before you publish them, and its Makers edition's "data anchor" rule is the same rule as our source ledger: if you can't trace it, it doesn't exist yet. This class makes that a machine that runs every time.
• Data Activism turns a public dataset into civic pressure. If your research question needs statistics rather than sources, its Method Selection Cheat Sheet is where you go.
• The school's Claim Autopsy bench takes one claim apart (source, method, alternative explanations, verdict). Use it in Block 2 on the claim your decision hangs on. The Research Synthesis bench walks ULM's three-paper synthesis; Block 4's Card 14 is its work version.
• Context Engineering and RAG & Memory teach how retrieval works underneath. This class teaches a non-engineer to investigate: assignment, provenance, independence, uncertainty.

This replaces Using Large Language Models

Using Large Language Models taught people to ask better. Its research material lives on here, pointed at a real job:

One correction carried over on purpose: ULM 3.2 quotes arXiv submission counts and "AI contributed to" fractions with no source. We don't reuse those numbers, and Block 2 uses that page as an audit exercise for the facilitator.

Handouts

• Prompt Cards: every prompt of the day, in order, with the spec-card slot it fills, including the Research Assignment template
• Practice Pack: the endangerment-finding repeal, with fourteen verified sources, five real disagreements and a timeline, as of 2026-11-10
• Spec Card: the shared one-page machine description

Last updated: 2026-11-10. Product features are as of Nov 2026. Facilitator: re-check Projects, deep research / Research mode, scheduled tasks and NotebookLM before each run.

↑ Contents

Block 1: Turn a Question into an Assignment

Time: 10:00–10:50 · Newsroom: the assignment desk · Atlas: /x/ai-at-work-research-map#d-assign

What you'll have at the end of this block

• A Project (ChatGPT or Claude) called Research Desk: <your topic>, with your Spec Card at the top of its instructions.
• A Research Assignment in those instructions that names the decision it serves, who makes it and by when, its scope, its source rules, and what would change your mind, all written before a single search.
• A model-fit note: which mode does this job (plain chat, web search, or deep research), with the evidence from your own three-way test.

ULM foundations reused here

• 1.4 LLM Assistant Onboarding Process: the flipped interaction. Instead of telling the AI what you want, get it to interview you.
• 1.2 LLM Model Comparison and Selection: picking a model, now for one workload and on evidence you gathered.
• 2.1 LLM Critical Thinking Vocabulary: "better" means nothing. "What would change our mind" is the research version of naming the criterion.

The idea

Here's what most people type:

What's going on with the EPA endangerment finding?

And they get back four tidy paragraphs and some links. It reads fine. Nobody can tell you if it's any good, because nobody said what it was for. "Good" isn't a quality — it's a sound you make with your face.

Now here's what a newsroom's assignment desk would send out:

Dana decides by January 15 whether our 2027 climate plan models transport emissions with or without federal vehicle standards. Find out whether those standards are likely to come back before 2027, using primary documents first. We'd change our answer if the D.C. Circuit sets a schedule or rules, or if the EPA acts on the reconsideration petition.

Same topic. But now the research has a reader (Dana), a deadline, a scope (vehicle standards, 2027), a source rule (primaries first), and a stopping condition (it changes our answer if X). Every one of those tells the machine what to throw away, and most of research is throwing things away.

That's the whole block, really. A question is what you're curious about. An assignment is what somebody has to decide. The desk refuses to work on the first one.

Why the decision goes first

Because it decides everything downstream. Which sources count: Dana needs the court docket, not op-eds. How confident you need to be: "with or without" can be both, as scenarios, if the answer is uncertain. When to stop: once the answer can't change the decision, more research is a hobby. If you can't name who decides what, by when, you don't have a research problem yet. You have a curiosity, and that's fine, but don't build a department for it.

Who's in the room, and who isn't

Your assignment has a decision-maker. It also has people the decision lands on who won't see your brief. Write them down now, in the assignment. You'll need them in Block 4, and if you don't name them here, your sources won't either. (Read the "who's behind it" column of the Practice Pack: agencies, lawyers, scientists, advocates, trade press, a blog. Nobody who lives beside the freeway.)

Worked example: the practice pack

Rough question: "Is the endangerment finding repeal going to stick?"

After the flipped interview (Card 2), the Project asked seven questions. Three of them changed the assignment:

• "Who makes the decision this research feeds?" → the sustainability director, not "the city."
• "What would they do differently depending on the answer?" → model transport emissions with standards, without, or both. That turned a yes/no forecast into a scenario question.
• "Who's affected but won't read the brief?" → residents on the freight corridors where truck emissions concentrate, and drivers of older cars, who'll face fuel costs either way.

The assignment it produced:

RESEARCH ASSIGNMENT
Question:            Will federal vehicle GHG standards be restored before our 2027 climate
                     plan update, and what should the plan assume meanwhile?
Decision it serves:  The sustainability director decides by Jan 15 whether the 2027 plan
                     models transport emissions WITH federal standards, WITHOUT, or BOTH as
                     scenarios.
Scope:               in: the EPA's 2026 rescission (91 FR 7686), the D.C. Circuit challenges
                     (No. 26-1037 and the states' petition), the EPA reconsideration
                     petition, any congressional action. out: the climate science itself
                     (the rule says it doesn't rest on it), state-level rules, other countries.
                     window: 2025-08-01 (proposal) to today. geography: federal.
Source rules:        primary first (Federal Register, court filings, the agency's own
                     releases); every claim with its quoted sentence; count origins; say who
                     is behind each source; name whose voice is missing.
What would change
our mind:            a D.C. Circuit briefing schedule, argument date, stay or ruling; EPA
                     acting on the 2026-04-16 reconsideration petition; legislation.
Deadline:            Dec 15 (brief), so the director has a month.
Format:              one-page brief: claim / confidence / source; what we don't know; whose
                     voice is missing; what would change this.

Notice the scope says out: the climate science itself, with a reason. The final rule says the EPA "is not basing this action on a new finding" and rests on a legal reading of the Clean Air Act (91 FR 7688). For this decision, a pile of climate-science sources would be on-topic and useless. That's the assignment doing its job: it threw away a whole genre before anyone searched.

Exercise: build the assignment desk

1. Make the Project (5 min). ChatGPT: Projects → New project. Claude: Projects → Create project. Name it Research Desk: <topic>. (As of Nov 2026. Facilitator: check the menus haven't moved.)

2. Draft the assignment by hand (8 min). Copy the Research Assignment template from Prompt Cards Card 1 into a note and fill in every line yourself. If a line is blank, leave it blank. Don't make it up.

3. Get interviewed (10 min). In a chat inside the Project, paste:

I'm about to give you a research assignment, but I don't trust my own first draft of it.
Before you research anything, interview me. Ask me one question at a time — at least seven —
about the decision this research serves, who makes it, what they already believe, what
deadline they're on, what would change their mind, which sources they'd dismiss on sight,
and who is affected by the decision but won't be in the room.
After the last answer, rewrite my assignment using the RESEARCH ASSIGNMENT template.
Do not search the web yet.

Here's my draft:
<paste your draft>

Answer honestly. "Dana thinks the lawsuit will win because her old boss is on it" is more useful than "stakeholders have varied views." Use voice input if you can, because you'll say more.

4. Write the spec card into the instructions (10 min). Paste Card 4 (the spec card), then the finished assignment, then Card 3 (the source rules) into the Project instructions. Fill the Gate slot with a real name: yours, or the person who signs off. "Someone reviews it" is how nobody reviews it.

5. The model-fit test (12 min). Paste Card 5 three times: plain chat with search off, chat with web search on, and deep research (ChatGPT) or Research mode (Claude). For each one, open the five links and count two things: how many quoted sentences you actually found on the page, and how many sources were primary. Write the three pairs of numbers in your spec card under a line Model fit:, with your choice and one reason. If you have NotebookLM, note that too: it can't search the open web the way deep research does, but it's the best auditor you'll meet in Block 2.

Using my RESEARCH ASSIGNMENT, give me only the five most important claims, each with one
source and the quoted supporting sentence. Mark each source primary or secondhand.
Stop at five.

The check

The assignment names its decision. Your partner reads only the "Decision it serves" line and says back: who decides what, by when. If they can't say all three, it isn't done.

Then a second, harder one: your partner reads your "What would change our mind" list and names one item that wouldn't actually change the decision. If they find one, cut it.

Common failures, and what they look like

• A topic dressed as a decision. "Decision it serves: understanding the landscape." Nobody decides "understanding." Ask: what will somebody do differently?
• A decision nobody owns. "The organisation will decide…" Organisations don't decide. People in them do. Name the person.
• "What would change our mind: new information." That's everything. Name the event: the ruling, the filing, the number crossing a line.
• Scoping in the thing you already believe. If every source rule happens to favour the answer you want, the critics will catch it in Block 4. Better to catch it now.
• Skipping the interview because you "already know." You know the summary. The interview pulls out the details, and the details are the assignment.
• Model fit by vibes. "Claude felt more thorough." Count the sentences you found on the page. That's the measure.

Facilitator note

• Timing: Project setup 5 · hand draft 8 · interview 10 · instructions 10 · model test 12 · check 5. The model test runs long because deep research is slow (often many minutes, as of Nov 2026). Have people start the deep-research run first and do the plain-chat and search runs while it works.
• Before the first feature: ask the room, "What does your Monday actually look like — what do you have to find out, for whom?" Get three answers out loud before anyone opens a Project. The students are the authorities on their work. Start there.
• Watch for: people writing the question and skipping the decision. Walk the room and read only the "Decision it serves" line. Also watch for anyone whose plain-chat run returned confident links: have them open one now. A fabricated citation in the first hour is the best thing that can happen to a room.
• Re-check before class: where Projects, deep research and Research mode live in each product, and what the free tiers include. Note it at the top of the room's shared notes.
• Think-pair-share: "Read your partner's assignment. What's one source they'd find that's on-topic and useless for their decision?" Two minutes in pairs, then take three answers. On the practice pack, the answer is "the climate science," and the reason is in the rule itself.
• Atlas: project /x/ai-at-work-research-map#d-assign and do its bench (which of four assignments names a decision?) as a room before the check.

Standards claimed: cap.produce-a-research-assignment-that-names-the-decision-it-serves (assessed), cap.produce-a-spec-card-for-a-recurring-task (assessed), cap.produce-a-reusable-system-prompt (assessed), cap.follow-a-model-quickstart (practiced), cap.classify-models-by-fit-for-a-workload (practiced).

↑ Contents

Block 2: Gather Sources and Audit Every Citation

Time: 11:00–11:50 · Newsroom: the reporters and the fact-check desk · Atlas: /x/ai-at-work-research-map#d-gather

What you'll have at the end of this block

• A source ledger from one real deep-research run on your assignment: at least 15 rows, every row with the same fixed fields.
• A citation audit of at least ten of those citations, each one opened by you and marked holds / wrong claim / dead / secondhand, with the supporting sentence quoted where it holds.
• At least one new source rule in your Project instructions, written from a failure your audit caught.
• Your load-bearing claim located in its primary source, page or section named.

ULM foundations reused here

• Exercise 1.3 Web Search and Fact-Checking, upgraded. ULM checked that the AI gave links. Here you open ten and find the sentence.
• 2.4 LLM Projects and Context Management: the ledger lives in the Project, so every later block reads from the same rows.
• 3.2 LLM Research Paper Synthesis's "fact-check list pattern," with one change: the model doesn't get to grade its own homework. You do.

The idea

A deep-research run hands you twenty links and a confident summary. It looks like research. It's a reporter's notebook, and a newsroom never prints from a notebook. It sends the notebook to the fact-check desk, where someone opens every source and finds the sentence.

A citation fails in four ways, and you'll see all four today:

The dangerous one isn't dead. Dead links are easy to catch, because they look broken. The dangerous one is wrong claim: a real, official, relevant page that says something next to what the summary says. It survives a glance, and a glance is all most people give it. That's why the audit is done with eyes on the sentence, not on the link.

A blocked page isn't dead. Paywalls and geo-blocks exist (the Bangor Daily News blocks some visitors from California, for one). Write "couldn't open," find another copy of the same story or the primary, and audit that.

Why it gets these wrong

Because a model writing a summary predicts what a citation for that sentence would plausibly look like, and plausible isn't the same as checked. Deep-research modes are far better at this than plain chat. You saw that in Block 1's test. But "better" isn't "done," and the failure that remains is the subtle one: right document family, wrong version. Which is exactly what happens in the practice pack.

Worked example: the practice pack audit

Below are seven ledger rows as a deep-research run might return them. The rows are illustrative. Every source, URL and quoted sentence is real and was checked on 2026-11-10.

Tally: 2 holds, 1 holds-but-out-of-scope, 1 wrong claim, 2 secondhand, 1 dead. That tally is the "one failure the check caught" box on your portfolio.

The source rule row 2 produced, pasted into the Project instructions:

8. Proposed rule ≠ final rule. Never cite a proposed rule, a draft or an earlier version for
   what the final version says. Check the date and the document's status on every row.

The locate-it exercise on row 4 is the one to remember. The secondhand number wasn't false, exactly. But the primary was more specific than the claim made from it, and the specific version tells a different story. That's what the "locate" standard asks you to find: one case where the source is weaker than (or different from) the claim built on it.

Exercise: the reporters and the fact-check desk

1. Send out the reporters (start immediately; it runs while you do step 2). Inside your Project, with deep research (ChatGPT) or Research (Claude) on:

Run my RESEARCH ASSIGNMENT under my SOURCE RULES.
Output a SOURCE LEDGER table with exactly the fields in my spec card, at least 15 rows.
Then list:
- claims you could not find a primary source for
- any two sources that disagree, with both sentences quoted
- whose voice is missing from the ledger
Do not write a conclusion yet.

If it comes back with different fields than your spec card, don't fix the table by hand. Change the prompt ("Your table has 7 fields; my spec card has 9. Redo it with exactly these fields: …") and run it again. Getting the output you specified by changing the prompt, not the spec, is the skill.

2. Set up the audit sheet (while it runs). Make a copy of the ledger columns plus two more: Audit and What I found. NotebookLM option: once you have the ledger, add the primary documents as sources in a notebook. When you ask it about a claim, its citations jump to the passage. Use it to find the sentence. You still mark the row.

3. Audit ten (25 min). Pick ten rows, including the three the model said it was least sure of, and the one your decision hangs on. For each: open the URL yourself, find the sentence, mark it. Don't ask the model whether its own citation holds. That's not an audit, that's a second opinion from the defendant.

4. Feed the failures back (5 min):

Here's my audit of your ledger: <paste rows with your marks>.
For every row marked wrong claim, dead or secondhand: find the primary source and the exact
supporting sentence, or delete the claim. Don't soften a claim to keep a row alive.
Then add this rule to the top of your answer and follow it in every future run:
"<the specific failure you caught, as a rule>"

Copy the new rule into the SOURCE RULES in your Project instructions.

5. Locate the load-bearing claim (8 min). For the one claim your decision hangs on:

This claim is load-bearing: "<claim>". Find the primary document it rests on. Give me the
section, page or paragraph number and quote the sentence. Then tell me whether the primary
says MORE, LESS or SOMETHING DIFFERENT than the claim as written — be specific.

Then open the primary and check its answer. Use the school's Claim Autopsy bench on that one claim if you want to take it fully apart.

The check

The citation audit. At least ten citations from this run, each marked holds / wrong claim / dead / secondhand, with the quoted sentence where it holds. Put the tally at the top: n holds · n wrong claim · n dead · n secondhand. Your partner picks two of your "holds" rows at random and finds the sentence on the page themselves in under two minutes each. If they can't, the mark changes.

Common failures, and what they look like

• Fabricated citations. A report title that sounds exactly right, a plausible URL, no such page. Mostly from plain chat, rarer in deep research, not zero.
• A real link that doesn't say the thing. The commonest and the costliest. On the practice pack: the proposal cited for the final rule.
• Secondhand-of-secondhand. A blog quoting a newsletter quoting the AP quoting a press release. Each hop rounds the number and drops a qualifier ("nearly 600,000" from "approximately 572,000 … from more than 31,000 unique entities").
• Auditing the summary instead of the source. Reading the model's paragraph and nodding. Open the page.
• Marking "dead" when it's blocked. Paywalls and geo-blocks aren't dead links. Find another copy.
• Keeping out-of-scope rows because they hold. True and irrelevant still goes. Your assignment's scope is a source rule too.
• Letting the model re-audit itself. It'll find its own citations convincing.

Facilitator note

• Timing: start the run 2 · audit sheet setup 5 (while it runs) · audit 25 · feedback prompt 5 · locate 8 · partner check 5. Deep research can take a long time (as of Nov 2026); if a run is still going at minute 10, have that student audit the practice-pack ledger rows above while they wait.
• Watch for: people auditing by hovering over links. Insist on the sentence. Also watch for despair when the tally is bad. A bad tally is the point. Say it plainly: you just found what would've gone out under your name.
• ULM correction: ULM's 3.2 LLM Research Paper Synthesis states arXiv submission counts and "AI contributed to" fractions with no source. If the room needs a second audit target, use those sentences: can anyone find a primary for them?
• Re-check before class: that the practice-pack URLs still resolve (row 5's 404 is part of the exercise; if WTOP restores that URL, swap in another dead one from your own run).
• Think-pair-share: "Which of your marks would your boss have caught? Which would nobody have caught until it was quoted back at you?" Pairs, two minutes. The answer is usually "wrong claim," and that's the lesson.
• Atlas: project /x/ai-at-work-research-map#d-gather. Its bench is five practice-pack citations to mark, with feedback on each. Do it as a room before step 3.

Standards claimed: cap.verify-a-research-tools-citations-say-what-it-claims (assessed), cap.locate-a-claim-in-its-primary-source (assessed), cap.characterise-a-models-failure-modes (practiced: the tally by failure kind, and the source rule each one produced), cap.elicit-output-from-a-model (practiced: the fixed-field ledger got by changing the prompt).

↑ Contents

Block 3: Count Your Sources and Map the Disagreement

Time: 12:00–12:50 · Newsroom: the debate room · Atlas: /x/ai-at-work-research-map#d-compare

What you'll have at the end of this block

• An independence count for the finding your decision leans on: how many links support it, how many independent origins those links come from, and which origins have a stake in the outcome.
• A contradiction table in your Project, fields fixed: where two sources disagree, both sentences quoted, what kind of disagreement it is, and what would settle it.
• An uncertainty map: every claim sorted into known, contested and unknown, with the unknown that matters most to the decision named.
• One contradiction resolved, or left open on purpose, with what would close it.

ULM foundations reused here

• 2.1 LLM Critical Thinking Vocabulary: "significance" and "accuracy" as named criteria, applied to how many sources and which kind, not "is it reliable?"
• Exercise 1.3 Web Search and Fact-Checking's "do multiple sources agree?", corrected: agreement means nothing until you know whether they're the same source.
• Exercise 2.4 Data Extraction and Structured Information: fields fixed before you read, so the table can't describe the sources back to you.

The idea

Open the Practice Pack and look at source S9. It's a story called "EPA sued over rescission of greenhouse gas endangerment finding." You can find it on ESG Dive. And on Manufacturing Dive. And on WardsAuto. Three outlets, three URLs, three logos. Same reporter (Jeff Kinney), same publication time down to the minute (2026-02-20, 10:05), same paragraphs. One publisher with three mastheads.

Now read what it says. Its sharpest lines are quotes: Environmental Defense Fund's counsel "said in a statement," the American Public Health Association's CEO "said." Those are parties to the lawsuit. So the claim "the repeal is illegal and dangerous" rests, in all three links, on one origin: the challengers' own statements. That's fine to cite as what the challengers say. It's not three independent confirmations of anything.

That's the independence collapse. Count origins, not links. Ten articles quoting one press release are one source, and that source has an interest.

The same thing happens with the AP. The story about the health groups' lawsuit (S8) runs in the News Tribune, TribLIVE, the Bangor Daily News and hundreds more. Deep research loves wire copy, because it's everywhere, so it looks corroborated. It's one newsroom's reporting.

Not every disagreement is a disagreement

When two sources say different things, figure out what kind of different before you pick a side. In the practice pack, they come in four kinds:

Most "contradictions" in a research run are the first three. Those you can resolve. The fourth kind is the one your brief has to be honest about.

The power question, stated directly

Who's behind each origin? In the practice pack the loudest origins are an agency defending its own rule, the parties suing it, a law firm writing for its clients, a law-school centre that supports state attorneys general, and a blog that opposes the finding in its first paragraph. None of that makes them wrong. It tells you what each would never say, and who's not in the list at all. Write the "who's behind it" field as fact, not insult: "a party to the case" is information. "Biased" isn't.

Worked example: the practice pack

Independence count for the finding "the challengers argue the repeal is unlawful":

Six links. Two newsrooms. One origin for the claim itself: the challengers. Every origin is a party. For this finding that's correct, because it's a claim about what they argue, and the petition is the primary. If the ledger had used these six links to support "the repeal is unlawful," the collapse would show that nobody independent of the case said so.

One contradiction, worked through. How many governments sued in March?

• S10 (AP): "In all, 24 states, 10 cities and five counties joined the lawsuit." The same story then names nine cities.
• S11 (NYU's centre): "a coalition of 25 attorneys general, 12 cities and counties, and the Governor of Pennsylvania."
• Counting the AP's own list gives 23 state attorneys general plus D.C. and the U.S. Virgin Islands, which is 25 attorneys general. Pennsylvania joined through its environment department, which makes 24 states.
• Kind: different definitions (attorneys general vs states). Status: the states/AGs part is resolved. The cities count is open until someone reads the caption of the petition itself, so the brief says "about two dozen states and roughly a dozen cities and counties (counts differ by source; see ledger)" and doesn't pick one.

Uncertainty map (excerpt):

The unknown that matters most to the decision: the court's timeline. If there's no ruling before 2027, the plan can't count on federal standards, whoever is right on the law. That's what pushes the recommendation toward "model both as scenarios."

Exercise: the debate room

1. Count the origins (12 min). Pick the finding your decision leans on most. In your Project:

For the finding "<finding>", list every source in the ledger that supports it. For each one,
trace the claim back to where it ORIGINATED: the press release, the filing, the wire story,
the dataset. Group sources that share an origin. Then tell me:
- how many links support it
- how many independent origins support it
- whether any origin is a party with a stake in the outcome

Then check two of its "same origin" groupings yourself: same byline? same timestamp? same quotes? Paste one sentence into a search engine in quotation marks and see how many places it turns up.

2. Build the contradiction table (15 min):

Build a CONTRADICTION TABLE from the ledger. Fields fixed:
point of disagreement | source A + quoted sentence | source B + quoted sentence |
kind (different events / different definitions / different time / genuine conflict) |
what would resolve it | status (resolved / open)
Don't resolve anything by majority vote. Don't resolve anything by trusting the more
famous source. If it's open, say what evidence would close it.

Audit every quoted sentence in it, the same way you did in Block 2. Contradiction tables are where invented quotes like to hide, because a disagreement looks more interesting with a sharp quote on each side.

3. Resolve one, or leave it open on purpose (8 min). Pick one row. Go to the primary. Either resolve it (and write how, in one line) or mark it open and write what would close it and when you might know.

4. Map the uncertainty (10 min):

Sort every claim in the ledger into three bins:
KNOWN — a primary source says it, and nobody credible disputes it
CONTESTED — credible sources disagree; name them
UNKNOWN — nobody can know yet; name the event that would tell us, and its likely date if any
Then tell me which UNKNOWN matters most to the decision in my assignment, and why.

Keep the UNKNOWN list. It becomes your watch list in Block 6.

The check

The independence collapse: your partner reads your independence count and picks one "independent" origin to challenge: show them it doesn't trace back to another one in the list. If you can't, it stays. If you can, the count drops.

One contradiction resolved or explicitly left open, in the table, with the kind named and the evidence that settled it (or would).

Common failures, and what they look like

• Counting links as sources. "Twelve sources confirm…" when nine are one wire story.
• False balance. Two sides get equal rows because there are two sides. If one side has the primary document and the other has a blog post, the table should show that, not a tie.
• Resolving by reputation. "The National Academies outrank a blog, so the blog is wrong." Maybe, but on which claim? The blog's date was a different event, not a wrong one.
• Resolving by majority. More sources saying a number doesn't make it right if they all copied it.
• Invented disagreements. A model asked for contradictions will find some. Check that both quoted sentences exist and that they're actually about the same thing.
• An uncertainty map with nothing in UNKNOWN. Everything live has an unknown. If yours is empty, you've hidden it in CONTESTED.

Facilitator note

• Timing: independence count 12 · contradiction table 15 · resolve one 8 · uncertainty map 10 · check 5. Lunch is right after this block, so stop at 12:50 even if the map is rough; it gets used again at 2:30.
• Watch for: students treating "who's behind it" as an accusation. Model the flat version out loud: "A party to the case. A law firm writing for clients. Self-funded by the Academy." It's information, not a verdict. Also watch for students who decide the question here. The debate room doesn't print anything.
• The power question: ask the room directly, "Whose voice isn't in any of these origins?" On the practice pack: people who live beside freight corridors, drivers of older cars, transit riders. Don't preach it. Ask it, then wait.
• Re-check before class: the D.C. Circuit docket for No. 26-1037. If there's news, the uncertainty map's UNKNOWN row changes, and that's a gift for Block 6.
• Think-pair-share: "Take your contradiction table's first row. Which of the four kinds is it? Convince your partner." Two minutes, then three pairs report. Most will turn out to be different events or different definitions, and the room learns that from each other.
• Atlas: project /x/ai-at-work-research-map#d-compare. Its bench is the independence collapse on six real practice-pack links. Do it before step 1.

Standard claimed: cap.characterise-whether-sources-are-independent (practiced).

↑ Contents

Block 4: Send It to the Critics

Time: 1:30–2:20 · Newsroom: the critics' table · Atlas: /x/ai-at-work-research-map#d-critics

What you'll have at the end of this block

• A first-draft position on your question, written by synthesising the primary sources in your ledger (not the secondhand ones).
• Three critic personas in your Project instructions: the skeptical domain expert, opposing counsel, and the person most affected, each specific to your field.
• Your ruling on every objection they raised: accept or reject, with a one-line reason.
• A critic diff: every sentence that changed, before and after, and the objection it answers. Nothing else changed.

ULM foundations reused here

• 2.2 LLM Expert Business Analysis and Exercise 2.1 Multi-Perspective Business Analysis: the bell hooks / Emma Goldman pass. Same move, pointed at a research draft and at power: whose costs aren't counted, and who the evidence leaves out.
• Exercise 1.4 Business Planning with Expert Personas: expert framing, here with a stopping condition so the critics don't just perform.
• 3.2 LLM Research Paper Synthesis and Exercise 3.2 Research Paper Analysis: digest each paper, synthesise from the digests, then fact-check the synthesis.
• 2.1 LLM Critical Thinking Vocabulary: objections must name a criterion (accuracy, relevance, significance), never "make it stronger."

The idea

In ULM you asked a business analyst for ideas, then brought in bell hooks and Emma Goldman, and the plan changed: "Don't scale. Scale is where capitalism creeps in." That pass works because each persona has a different idea of what counts as a cost.

A research draft needs the same thing, with sharper briefs. A newsroom's critics' table reads the draft as the people it will make angry, before it runs. Three seats:

1. The skeptical domain expert. Not "an expert": a specific one. For the practice pack, an administrative-law professor who has read all 111 pages of 91 FR 7686. Catches misread mechanics, numbers compared that don't measure the same thing, and definitions that slid between paragraphs.
2. Opposing counsel. Argues the strongest honest case for the outcome your draft treats as less likely. Catches the assumptions that happen to match what you wanted.
3. The person most affected. A specific person the decision lands on, someone who won't read your brief. For the practice pack: a person who lives beside a freight corridor and drives a 2011 car. Catches whose costs aren't counted, whose voice no source in the ledger carries, and what your neutral sentence sounds like to someone it happens to.

The rule that makes this work: the critics object, and you rule. The model doesn't get to revise the draft until you've accepted or rejected each objection yourself. Otherwise you get a draft that's been sanded smooth by three imaginary people, and you can't tell which changes were right. The diff is the evidence that a human made the calls.

Expert framing has a limit

A persona is a prompt pattern, not a person. It'll sound like opposing counsel without knowing the case law, and the person-most-affected seat can turn into a costume. So every objection has to point at a sentence in your draft and a row in your ledger, or say NO SOURCE. An objection with NO SOURCE isn't wrong. It's a question you now have to research or flag as open. Write down where the pattern stopped helping. You'll need that note when you reuse these critics on a different desk in Block 6.

Worked example: the practice pack

Draft (first two sentences, written from primary digests with Card 14):

The EPA's repeal of federal vehicle greenhouse-gas standards has been in effect since 2026-04-20 and is being challenged in the D.C. Circuit. Because the repeal is likely to be overturned, the 2027 plan should continue to assume federal standards, which save households $1.4 trillion.

Objections (illustrative, of the kind the three critics raise):

The critic diff:

- Because the repeal is likely to be overturned, the 2027 plan should continue to assume
-   federal standards, which save households $1.4 trillion.
+ No court has ruled or set a schedule (as of 2026-11-10), and the only sources predicting
+   the outcome are parties to the case, so the 2027 plan should model transport emissions
+   both WITH and WITHOUT federal standards.               [opposing counsel, accepted]
+ (Cost figures in the record measure different things and aren't used here; see ledger.)
+                                                           [skeptical expert, accepted]

Sentence 1 didn't change. That's in the diff too: rejected objections are part of the record.

Exercise: the critics' table

1. Synthesise from the primaries (12 min). This is ULM 3.2's three steps (digest, synthesise, fact-check) on your ledger:

For each PRIMARY source in the ledger, write a 120-word digest that another model could use
to find where the sources agree and disagree: what it claims, on what basis, what it
explicitly does NOT claim, and its date. Then, using only those digests, state the
strongest position the evidence supports on my question, name where two primaries
disagree, and say which reading you took and why. Then fact-check your own position
against the ledger and mark any sentence with no ledger row as UNSUPPORTED.

Delete or rewrite every UNSUPPORTED sentence before the critics see it. (The Research Synthesis bench walks the same three steps on papers, if you want the long version.)

2. Seat the critics (8 min). Add this to your Project instructions, filling in the specifics for your field. Specific beats famous: "a procurement officer at a mid-size county who's been burned by a vendor before" beats "a business expert."

CRITICS
1. The skeptical domain expert: <specific>. Catches: misread mechanics,
   numbers compared that don't measure the same thing, definitions that slid.
2. Opposing counsel: argues the strongest honest case for the outcome my draft treats as
   less likely. Catches: assumptions that happen to match what I want.
3. The person most affected: <a specific person the decision lands on>. Catches: whose
   costs aren't counted, whose voice the sources leave out, what the brief would sound like
   to them.
Each critic: raise at most 3 objections, each tied to a specific sentence in the draft and
to a source in the ledger (or flagged NO SOURCE). No general advice. No praise.

3. Convene them (5 min):

Here's my draft brief: <paste>. Convene the CRITICS. Each critic gives their objections
separately. Then STOP. Don't revise the draft. I'll rule on each objection myself.

4. Rule (12 min). For every objection: ACCEPT or REJECT, one line why, checked against your ledger and your source rules. Open the source if the objection cites one you haven't audited. You're the editor. Disagreeing with a critic is allowed, and rejecting a bad objection is as much evidence of judgment as accepting a good one.

5. Get the diff (5 min):

My rulings: <objection 1: ACCEPT / REJECT + one-line reason> ...
Revise the draft for the ACCEPTED objections only. Show me a diff: every sentence you
changed, before and after, and which objection it answers. Change nothing else.

Check the diff against your rulings. If it changed a sentence nobody objected to, reject that change too.

The check

The critic diff. Your partner reads your rulings and the diff side by side and checks two things: every changed sentence traces to an accepted objection, and at least one objection was rejected with a reason. (If you accepted everything, you weren't editing, you were obeying.)

Common failures, and what they look like

• Critics who all sound the same. Three polite reviewers who each suggest "more nuance." Make the briefs specific, and cap them at three objections each.
• Critics with no sources. "Opposing counsel would argue…" with nothing behind it. Every objection points at a ledger row or says NO SOURCE.
• Letting the model revise before you rule. You get a smoother draft and no idea which changes were right.
• False balance as a fix. Opposing counsel's objection gets "accepted" by adding a both-sides sentence with no evidence on the other side. Accept the objection, then follow the evidence, which may mean the draft stays where it was.
• The person most affected as a costume. A paragraph of invented feelings in a made-up voice. The useful version names a cost, a gap in the sources, or a missing voice, and then you go find that voice for real.
• Synthesising from secondhand sources. If the digests are of news stories, the synthesis is of news stories. Primaries only.

Facilitator note

• Timing: synthesis 12 · seat the critics 8 · convene 5 · rule 12 · diff 5 · partner check 8. People back from lunch are slow for ten minutes. Start with the think-pair-share below, not with typing.
• Teaching to transgress: the person-most-affected seat is where this class says out loud that research has power in it: who gets studied, who gets quoted, who gets counted. Say it once, directly, without a lecture: "Your sources decide whose problem this is. Who's missing from yours?" Then let them work.
• Watch for: people accepting every objection because the critic sounded authoritative. Ask, "Which ledger row backs that?" Also watch for anyone writing the critics as real named people's opinions. ULM's attribution rule still holds: it's AI reasoning in a persona, not a quote.
• Re-check before class: nothing product-specific here, apart from making sure Project instructions take the extra CRITICS section without truncation (long instructions have limits that change; as of Nov 2026 they're generous, but check).
• Think-pair-share (opener): "Name the person most affected by your decision who will never read your brief. What's one thing they know that none of your sources do?" Two minutes in pairs, take three answers.
• Atlas: project /x/ai-at-work-research-map#d-critics. Its bench is "which critic would catch this gap?" on four real practice-pack gaps.

Standards claimed: cap.generalise-a-reasoning-prompt-pattern (practiced: expert framing as three critic personas, reused unchanged in Block 6 with a note on where it stopped helping), cap.operate-a-model-as-a-first-pass-editor (practiced: every objection ruled on against your source rules, and only accepted ones changed).

↑ Contents

Block 5: Write the Brief and Run It Twice

Time: 2:30–3:20 · Newsroom: the front page · Atlas: /x/ai-at-work-research-map#d-brief

What you'll have at the end of this block

• The brief: one page, for the person who decides. The decision and the recommendation, a claim / confidence / source table, what we don't know, whose voice is missing, and what would change the recommendation. Signed by you.
• A second run of the same assignment (in the other product, or on a sibling question), and a diff of the two ledgers.
• A partner who has read your brief for sixty seconds and can say the decision and the biggest uncertainty back to you.

ULM foundations reused here

• 2.1 LLM Critical Thinking Vocabulary: the confidence labels are rubric levels with definitions, not moods.
• 3.2 LLM Research Paper Synthesis: the position you defend comes from the synthesis you built and the critics tested in Block 4.
• Exercise 1.3 Web Search and Fact-Checking's "compare with and without web search," grown up: compare two whole runs and read the differences.

The idea

A front page prints three things a notebook doesn't: what we're sure of, how sure, and what we don't know yet. Most AI research output prints only the first, in a confident voice, at the same volume for every sentence. Your reader can't tell the rule's publication date (certain, in the Federal Register) from a forecast of a court decision (a guess, from a blog).

So the brief has a fixed shape:

1. The decision and the recommendation, in two sentences.
2. Claim | confidence | source. Confidence has a definition, written into your instructions: - High: a primary source says it, and nobody credible disputes it. - Medium: a primary supports it but it's contested, or only good secondhand sources support it. - Low: a forecast, an inference, or a single interested source.
3. What we don't know, from the uncertainty map, each with what would tell us.
4. Whose voice is missing from the evidence.
5. What would change this recommendation, straight from the assignment.

No claim without a ledger row. That rule is what turns a summary into a brief.

Then run it again

A desk that worked once is a demo. Run the same assignment a second time (a week later, or in the other product, or on a sibling question) and diff the two ledgers. You'll find claims that appear in one run and not the other, the same claim cited to different sources, and occasionally the same claim with a different confidence. Every difference is something to look at. A claim that only one run found deserves a second audit. A claim both runs found from different origins just got stronger. A claim both found from the same origin didn't.

A second run that breaks something is the most useful thing that happens all day. You'd rather find the instability now than after the brief goes out.

Worked example: the practice pack brief

To: Sustainability director · From: · Date: · Signed:

Decision: whether the 2027 climate plan models transport emissions with or without federal vehicle GHG standards. Recommendation: model both as scenarios. As of 2026-11-10 no court has ruled or set a briefing schedule, so neither assumption is safe to plan on alone.

Claim	Confidence	Source
The EPA rescinded the 2009 endangerment finding and repealed all federal vehicle GHG standards; published 2026-02-18, effective 2026-04-20	High	S1, 91 FR 7686
The repeal rests on a legal reading of Clean Air Act §202(a), not a new scientific finding; the EPA says it is not relying on the DOE Climate Working Group report	High	S1, pp. 7688, 7691
Health and environmental groups (No. 26-1037, filed 2026-02-18) and about two dozen states plus cities and counties (filed 2026-03-19) are challenging it in the D.C. Circuit	High (filings) · counts differ by source	S6, S10, S11
No merits briefing schedule had been set as of mid-August 2026; an abeyance motion was reported filed 2026-07-29	Medium (one secondhand, opinionated source; not yet checked against the docket)	S13
A decision before the 2027 plan update	Low (forecasts only, from before the abeyance motion)	n/a

What we don't know: when the court rules (watch: a briefing schedule or argument date in No. 26-1037); whether the EPA acts on the 2026-04-16 reconsideration petition (watch: a Federal Register notice).

Whose voice is missing: every source is an agency, a court filing, a law firm, an advocacy group or the press. None is from residents of the freight corridors where truck emissions concentrate, or from drivers of older vehicles who carry fuel costs first. Next run's source rules add local health data.

What would change this: a ruling or a stay; EPA action on reconsideration; legislation. Any of those, and this brief gets re-run.

The partner test on this brief: "The decision is with-or-without federal standards in the plan. The biggest uncertainty is when the court rules." ✓

Run 2, diffed (illustrative of what a second run in the other product turns up):

Exercise: the front page

1. Write the brief (12 min). Add the confidence definitions above to your Project instructions, then:

Write the brief for <decision-maker>. One page. Structure fixed:
1. The decision and the recommendation in two sentences.
2. A table: claim | confidence (high / medium / low) | source (ledger row) — confidence is
   HIGH only if a primary source says it and it isn't contested.
3. What we don't know: the open items from the uncertainty map, each with what would tell us.
4. Whose voice is missing from this evidence.
5. What would change this recommendation.
No claim without a ledger row.

Read every row. Lower any confidence that's higher than its definition allows. Then sign it: put your name on it. You're the gate.

2. The partner test (6 min). Swap briefs. Your partner reads for sixty seconds, then closes it and says out loud: the decision, and the biggest uncertainty. If either is wrong or missing, the brief's first two sentences need work, not your partner.

3. Run it twice (start now; it runs during step 4). Run the gather prompt again (Card 6), either in the other product or on a sibling question that uses the same assignment and source rules. Practice-pack sibling: "What are the states arguing in their petition that the health groups aren't?"

4. Diff the runs (12 min):

Here are two SOURCE LEDGERS for the same assignment: RUN 1 <paste> and RUN 2 <paste>.
Tell me: claims in both · claims only in run 1 · claims only in run 2 · claims where the
two runs cite different sources or quote different sentences · any claim that changed
confidence. Don't merge them. I want the differences.

Audit every "only in one run" claim that's in your brief. Update the brief if anything moved, and note in the brief's footer that it was checked against two runs.

The check

The partner test (they name the decision and the biggest uncertainty after sixty seconds), and the run diff (two ledgers, the differences listed, and every one-run-only claim in the brief re-audited). Both go in the Project, where the portfolio link shows them.

Common failures, and what they look like

• Every claim is "high." The brief sounds sure of everything, including a court date nobody knows. Check each label against its definition.
• A conclusion that outruns the table. The recommendation says "likely overturned" and nothing in the table is better than Low on that. The critics should've caught it; the partner test catches it now.
• "What we don't know: further research is needed." That's not an unknown, it's a shrug. Name the event that would tell you.
• False balance on the front page. "Some say X, others say Y" when one side has the Federal Register and the other has a blog post. The confidence column is where you show the difference.
• Merging the runs. Asking the model to "combine the best of both" hides every difference you ran it twice to find.
• No signature. The brief goes out "from the desk." It doesn't. It goes out from you.

Facilitator note

• Timing: brief 12 · partner test 6 · second run starts (runs ~10+ in the background) · diff 12 · re-audit 8 · wrap 5. Start the second run before the partner test if deep research is slow that day.
• Watch for: briefs longer than a page. Make them cut. A director reads the first two sentences and the table. Also watch for students who skip the second run because the first "worked." The second run is the point.
• Re-check before class: the brief's worked example has dated facts. Update the table's "as of" line and the court status from the docket before projecting it.
• Think-pair-share: "What's one claim your second run found that your first didn't? Is it a new origin or the same origin again?" Pairs, two minutes. Take answers that show both cases.
• Atlas: project /x/ai-at-work-research-map#d-brief. Its bench is "set the confidence" on five practice-pack claims, each with feedback, so do it before step 1.

Standard claimed: cap.generalise-several-sources-into-a-position-you-can-defend (assessed: the brief takes one position, puts a ledger row behind every claim, and names one disagreement between two primaries with the reading you took and why. On the practice pack, that's the proposal vs the final rule on the DOE report).

↑ Contents

Block 6: Build a Second Desk and Set the Watch

Time: 3:20–4:00 · Newsroom: the beat · Atlas: /x/ai-at-work-research-map#d-watch

What you'll have at the end of this block

• A watch list on your first desk: a scheduled prompt that re-checks only the "what would change our mind" items and reports CHANGED or NO CHANGE, and ignores new articles that repeat an origin you already have.
• A second research desk in a new Project, for a different question in a different domain, built by you with no help from the facilitator: its own spec card, assignment, source rules and critics, and one real run with five citations audited.
• A short "what carried over, what changed" list comparing the two desks.
• Your portfolio link.

ULM foundations reused here

• 2.4 LLM Projects and Context Management: one Project per desk, so the instructions, ledger and history stay together.
• 1.4 LLM Assistant Onboarding Process: the flipped interaction again, this time comparing the new assignment to the old desk.
• 2.2 LLM Expert Business Analysis: the critics carry over, and you find out where they stop fitting.

The idea

A newspaper doesn't stop covering a story when it runs. A reporter keeps the beat: checks the docket, calls the clerk, reads the Federal Register every morning, and writes again only when something changed. Your desk can keep a beat too, if you tell it exactly what to watch and exactly what doesn't count as news.

What doesn't count as news is the important half. The practice pack is a good warning: one AP story became hundreds of URLs. A watch prompt that reports "new articles about the endangerment finding" will report the same story every week from a different masthead. Your watch prompt watches the events on your "what would change our mind" list, and reports a new article only if it carries a new origin.

Then you build another desk. That's the actual test of today. If the first desk was a set of prompts that happened to work on one question, the second one will be painful. If it was a machine, it'll be a new spec card, not a new skill. You'll find out which, and both answers are useful.

Worked example: the practice pack

The watch prompt for the endangerment desk:

WATCH: endangerment finding repeal. Check only for NEW evidence since <last run date> on:
1. D.C. Circuit No. 26-1037 (and the consolidated states' petition): briefing schedule,
   argument date, stay, abeyance ruling, or decision
2. EPA action on the 2026-04-16 reconsideration petition (Federal Register notice or EPA release)
3. Legislation restoring or barring federal vehicle GHG standards
For each: CHANGED (quote the new sentence, link, date) or NO CHANGE.
Ignore new articles that repeat an origin already in the ledger — list them under
"same origin, not new."
If anything changed, say whether it moves the recommendation in the brief. Draft only.

What a good watch report looks like (illustrative format):

1. No. 26-1037 — NO CHANGE since <date>. (Checked: docket listing.)
2. Reconsideration petition — NO CHANGE.
3. Legislation — NO CHANGE.
Same origin, not new: 4 reprints of the AP story from 2026-02-18; 1 blog post restating S13.
Recommendation: unchanged (model both scenarios).

Boring is the goal. A watch list that reports something every week is a news feed, and you'll stop reading it by the third week.

A second desk, sketched (a fundraiser's, from a different domain):

RESEARCH ASSIGNMENT
Question:            Does the <named> foundation's new strategy still fund small
                     organisations doing our kind of work?
Decision it serves:  Our development director decides by <date> whether to spend 30 hours
                     on their next open call.
Scope:               in: the foundation's own strategy page, guidelines, 990s, grants list
                     for the last 2 years. out: anyone's opinion of the foundation.
Source rules:        rules 1–8 from desk 1, plus: 9. A 990 or grants database beats a
                     press release for "who they actually fund."
What would change
our mind:            a grant under $X to an organisation our size in the last 2 years;
                     guidelines that exclude our category.

What carried over, what changed:

• Carried over unchanged: the spec card shape, the ledger fields, source rules 1–8, the audit marks, the brief format, the critic structure.
• Changed: what counts as primary (a 990 and a grants list instead of the Federal Register and court filings); the domain expert (a program officer who's read a hundred applications); the person most affected (the small organisations the strategy shift leaves out).
• Where the critics stopped helping: opposing counsel had little to argue with on a funding question. It became "the case that they still fund groups like ours," which was still useful, just not adversarial.

Exercise: the beat, then the second desk

1. Set the watch (8 min). Write your watch prompt from Card 17 in Prompt Cards, using your assignment's "what would change our mind" list. Schedule it (as of Nov 2026; facilitator re-checks):

• ChatGPT: run it as a Task: ask it to run the prompt weekly, or set it from the Tasks menu. Tasks notify you when they run.
• Claude: if your plan has scheduled tasks, set it weekly. If not, put a recurring calendar event titled "Research Desk: run WATCH" with the prompt pasted in the description. A rhythm you keep beats an automation you don't have.

Run it once now, to check it reports NO CHANGE correctly and lists same-origin articles separately.

2. Write the second spec card (8 min, solo, no help). A different question in a different domain from your own work. Use the Spec Card and Card 1 for the assignment. If you can't think of one, search regulations.gov for a rule in your field that's open for comment, and ask what your organisation would decide about it.

3. Stand up the second desk (10 min). New Project. Paste the new spec card and assignment. Then:

Here's the spec card and assignment for a new research desk. Before you research anything,
compare it to this working desk's instructions: <paste desk 1's instructions>.
List what carries over unchanged, what has to change for the new domain (source rules,
critics, what counts as primary), and anything in the new assignment that's still vague.
Ask me about the vague parts. Then wait.

Answer its questions, update the instructions, and run the gather prompt (Card 6).

4. Audit five (8 min). Five citations from the second desk's run, marked holds / wrong claim / dead / secondhand. Five, not ten, because you have eight minutes, and the first desk already proved you can do ten.

5. Write the carry-over list and submit (6 min). Three lines: carried over, had to change, where the critics stopped helping. Put it in the second desk's instructions under NOTES. Then share the link (custom GPT link, or the shared Claude Project or an artifact published from it) and check the five portfolio boxes in the syllabus.

The check

The five portfolio boxes, against your one link: spec card in the instructions · two real runs and their diff · one failure the audit caught and the source rule it added · the human gate named (who signs, and your critic rulings) · a second desk with a real run, plus the watch list scheduled on the first.

Common failures, and what they look like

• A watch prompt that watches a topic. "Tell me what's new about X" reports reprints forever. Watch named events.
• A second desk that's the first desk with the topic swapped. Same source rules, same critics, same "primary," copied blind. The carry-over list should have something in "had to change," or you didn't look.
• Picking a second question with no decision. End of day, tired, you'll want to research something fun. Fine, but then it's a hobby, not a desk. Name who decides what.
• Scheduling with no human reading it. The watch report drafts. Somebody reads it, and the spec card's Rhythm slot says who notices if it stops arriving.
• Skipping the audit on desk 2 because desk 1 passed. Different domain, different failures. A grants database fails differently from the Federal Register.

Facilitator note

• Timing: watch 8 · spec card 8 · second desk 10 · audit five 8 · carry-over and submit 6. This block is hands-off on purpose: don't help with desk 2 beyond pointing at the Prompt Cards. Help is what makes it not count.
• Watch for: students who finish the second desk in ten minutes. Ask to see their "had to change" line. If it's empty, ask what counts as a primary source in the new domain. Also watch for anyone whose watch list fires on the first run. Read it with them: is it a new origin or a reprint?
• Re-check before class: ChatGPT Tasks and any Claude scheduled-task feature: where they live, which plans have them, how many you can schedule (as of Nov 2026, these move often). Also the practice pack's docket status. If the D.C. Circuit has acted since the pack was written, the watch run comes back CHANGED, and that's a better demo than any slide.
• Think-pair-share: "Read your partner's carry-over list. What's one thing they said carried over that you think should have changed?" Two minutes, then close the day with three answers.
• Close the day by asking the room what their Monday looks like now. Same question you opened with. Listen for "I'll run the watch" and "I'll audit before I forward."
• Atlas: project /x/ai-at-work-research-map#d-watch. Its bench is "report or ignore?" on five possible watch-list hits. Do it before step 1.

Standard claimed: cap.generalise-a-working-ai-routine-to-a-second-case (assessed: the second desk, built solo, with its own spec card and real run, and the carry-over list naming what changed and what didn't). The critics reused here are the second half of Block 4's cap.generalise-a-reasoning-prompt-pattern evidence.

↑ Contents

Practice Pack: The Endangerment Finding Repeal

Bring your own question if you can. A question from your real work (a grant you're deciding whether to chase, a competitor you're deciding whether to answer, a rule your clients keep asking about) teaches you more than anything in this file, because you're the authority on whether the answer is any good. This pack is for when you can't bring yours: employer rules, privacy, or you just don't have one that's live this week.

It's a real question, live as of November 2026, with real primary sources that genuinely disagree with each other. That's the point of it. A practice question where every source agrees can't teach you to audit anything.

Facilitator: every URL below was opened and checked on 2026-11-10. The litigation is moving. Before each run of the class, re-check the D.C. Circuit docket for No. 26-1037 (CourtListener or PACER) and add anything new to the timeline. If the court has ruled, the pack gets better (it turns into a "what changed?" exercise for Block 6), not worse.

The question

On 2026-02-18 the EPA published a final rule rescinding the 2009 "endangerment finding" for greenhouse gases and repealing every federal greenhouse-gas emission standard for cars, trucks and engines. Two sets of petitioners are challenging it in the D.C. Circuit. Will federal vehicle greenhouse-gas standards come back before our next plan update, and what should our plan assume in the meantime?

The decision it serves (illustrative role)

You're a policy analyst in a city sustainability office. Your city's climate action plan is updated in 2027, and transportation is the biggest slice of its emissions inventory. Your director has to decide one thing: does the 2027 plan model transportation emissions with federal vehicle greenhouse-gas standards in place, without them, or both as scenarios? The answer changes how much the city has to do locally (charging, transit, zoning) to hit the same target.

That's an invented job on a real question. Swap in your own role if a different one fits you better: a fleet manager deciding on a 2027 vehicle order, a nonprofit deciding whether to file an amicus brief, a reporter deciding whether there's a story.

What would change our mind

Write this before you search (you'll do it in Block 1). For the practice pack, a reasonable version is:

• The D.C. Circuit sets a briefing schedule, holds oral argument, or rules in No. 26-1037.
• The EPA acts on the 2026-04-16 administrative petition asking it to reconsider.
• Congress passes anything that restores or bars vehicle GHG standards.
• A court stays the rule (pauses it while the case runs).

The source set

Every source here was opened on 2026-11-10. Type says whether it's the original document (primary) or someone reporting on one (secondhand). Who's behind it is the power question: who wrote it, who pays for it, and whose interest it serves. Look it up yourself before you trust our one line.

Whose voice is missing? Read the "who's behind it" column top to bottom. There are agencies, scientists, lawyers, advocacy groups, a trade-press publisher, a wire service and an opinion blog. There's nobody who breathes the air next to a freight corridor, drives for a living, or pays the fuel bill on an old car. They're the people the decision lands on. Your source rules in Block 1 should say what you'll do about that.

The disagreements already in this set

You'll find these yourself in Block 3. They're listed here so the facilitator can check that the room found them, and so you know the pack is worth the effort.

1. When did the repeal happen? The AP says the rule was finalized 2026-02-12 (the White House announcement). The Manhattan Contrarian says 2026-02-13. The Federal Register published it 2026-02-18, and it took effect 2026-04-20 (S1). Resolvable: they're four different events. Your ledger has to say which one a date refers to.
2. Did the EPA rely on the DOE report? The proposal (S2, footnote 10, p. 36292) says the DOE draft report "was reviewed and relied upon in formulating this proposal." The final rule (S1, footnote 8, p. 7691) says the EPA "is not relying on" either version of the report "for any aspect of this final action." Resolvable: both are true, about different documents. Anything written between August 2025 and February 2026 that says "the repeal relies on the DOE report" was describing the proposal. Cite it for the final rule and it's a wrong claim.
3. How many challengers? The AP (S10) says "24 states, 10 cities and five counties," then lists nine cities by name. NYU's center (S11) says "25 attorneys general, 12 cities and counties, and the Governor of Pennsylvania." Partly resolvable: the AP's own list has 23 state attorneys general plus D.C. and the U.S. Virgin Islands (25 attorneys general), plus Pennsylvania through its environment department (24 states). The cities number stays open until you read the caption of the petition itself. Leave it open in your brief rather than picking one.
4. Savings or costs? The AP (S8) reports that the White House and the EPA Administrator claimed $1.3 trillion in savings over three decades, while "a chart within the agency's regulatory analysis predicts $1.4 trillion in additional costs through 2055" from fuel and repairs. The final rule (S1, p. 7755) talks about "the estimated $1.09 trillion reduction in vehicle technology cost at a three percent discount rate." Not a contradiction yet: they measure different things (total savings, one category of added costs, one category of reduced costs) over different windows and discount rates. The mistake would be to put two of them side by side as if they answered the same question.
5. When will the court decide? A May 2026 post forecast oral argument in the fall and a decision by mid-2027. The Manhattan Contrarian (S13, 2026-08-13) reports no merits briefing schedule and an abeyance motion filed 2026-07-29. Open, and dated: the honest answer as of November 2026 is "we don't know, and here's what we're watching." That's a watch-list item (Block 6).

Timeline (as of 2026-11-10)

Run 2 and the second desk

• Run it twice (Block 5). Same assignment, a week later or in the other product (ChatGPT's deep research if you used Claude's Research the first time, or the other way round). Diff the two ledgers.
• A sibling question for the second run: "What are the states arguing in their petition (S10, S11) that the health groups (S6) aren't?" Same desk, same source rules, a different question.
• The second desk (Block 6). Pick a different domain from your own work. If you're stuck: a federal rule in your field that's open for comment right now. Search regulations.gov for your industry's agency and filter to "Open for Comment."

Why this question

It has everything a research desk has to handle: primary documents that are long and boring and public, a wire story syndicated everywhere (the independence lesson), the same trade-press story under three mastheads (the independence lesson again, harder to spot), a claim that was true of one document and false of the next (the citation audit), numbers that look like they contradict and don't (the uncertainty map), an outcome nobody knows yet (the watch list), and loud, well-funded voices on both sides with the affected public barely quoted (the critics' table).

It's politically charged. That's fine. The desk's job isn't to tell you what to think about climate policy. It's to make sure every claim in your brief is attached to a sentence somebody actually wrote, and that you know who wrote it.

↑ Contents

Prompt Cards: The Research Desk

Every prompt of the day, in the order you'll use it. Each card says which slot of the Spec Card it fills, so you can see the machine being built one part at a time.

Paste the bracketed parts with your real material. If you're on the practice pack, the brackets are filled in Practice Pack.

Product features change fast. Everything below about Projects, deep research, Research mode, NotebookLM and scheduled tasks is as of Nov 2026. Facilitator: click through each one the morning of class.

0. Where the desk lives

You build one Project and keep everything in it.

• ChatGPT: Projects → New project → name it Research Desk: <your topic> → Project instructions. You can also build it as a custom GPT (Explore GPTs → Create) if you want to share it as one link. Deep research is in the composer's tools menu. Scheduled re-runs are Tasks.
• Claude: Projects → Create project → "Set project instructions." Turn on Research (and web search) from the tools toggle in the chat. Scheduled re-runs depend on your plan and app; if you don't have them, a calendar reminder that says "re-run the watch prompt" is a fine rhythm.
• NotebookLM: not the desk itself, but the best place to audit — upload the primary documents as sources and every answer's citation jumps to the exact passage.

Card 1 · The Research Assignment template · slot: Intake

Fill this in by hand first. Every field. If you can't fill one, that's what you're missing, not what the AI is missing.

RESEARCH ASSIGNMENT
Question:            <one question, ending in a question mark>
Decision it serves:  <who decides what, by when — "Dana decides whether the 2027 plan models
                      transport with or without federal vehicle standards, by Jan 15">
Scope:               <in: ...   out: ...   time window: ...   geography: ...>
Source rules:        <what counts — see Card 3>
What would change
our mind:            <the specific evidence that would flip the answer — 2 to 4 items>
Deadline:            <date the brief is due>
Format:              <the brief: claim / confidence / source table + "what we don't know">

Card 2 · Flip the interview · slot: Intake (scoping)

From ULM's 1.4 LLM Assistant Onboarding Process: instead of telling it what to do, get it to interview you.

I'm about to give you a research assignment, but I don't trust my own first draft of it.
Before you research anything, interview me. Ask me one question at a time — at least seven —
about the decision this research serves, who makes it, what they already believe, what
deadline they're on, what would change their mind, which sources they'd dismiss on sight,
and who is affected by the decision but won't be in the room.
After the last answer, rewrite my assignment using the RESEARCH ASSIGNMENT template.
Do not search the web yet.

Here's my draft:
<paste your Card 1 draft>

Card 3 · Source rules · slot: Structure (what counts as evidence)

SOURCE RULES
1. Primary first. For every claim, find the document where it was first made: the rule, the
   filing, the dataset, the report. A news story about a document is secondhand.
2. Every claim gets: source title · publisher · URL · date published · date I accessed it ·
   primary or secondhand · the exact sentence that supports it, in quotation marks.
3. No sentence, no claim. If you can't quote the supporting sentence, write NOT FOUND
   and don't use the claim.
4. Count origins, not links. Ten articles quoting one press release are one source.
5. Say who's behind each source: who wrote it, who funds it, whose interest it serves.
6. Name whose voice is missing: who is affected by this decision and not quoted anywhere.
7. Dates are events. Say which event a date refers to (announced / signed / published /
   effective / filed / decided).

Card 4 · The spec card, inside the machine · slot: all seven

Paste this at the top of your Project instructions (ChatGPT: Project instructions or the custom GPT's Instructions box; Claude: Project instructions). Then paste Card 1 and Card 3 under it.

SPEC CARD
Intake:    a RESEARCH ASSIGNMENT (below). Refuse to research a bare question with no decision.
Structure: a SOURCE LEDGER, one row per claim, fields fixed:
           claim | source | publisher | URL | published (event) | accessed | primary/secondhand
           | who's behind it | supporting sentence (quoted)
           Never add, drop or rename a field.
Check:     I audit at least 10 citations by hand: holds / wrong claim / dead / secondhand.
           Tell me which 3 rows you are least sure of and why.
Gate:      <your name> signs the brief before it goes to <decision-maker>. You draft; you never send.
Act:       the brief goes to <where the decision gets made: the memo, the meeting, the doc>.
Ledger:    the source ledger + my audit + the diff between runs, kept in this Project.
Rhythm:    the watch prompt re-runs every <week/month>; it reports only what changed.

Card 5 · Model fit · slot: Structure (pick the tool)

From ULM's 1.2 LLM Model Comparison and Selection, pointed at research. Run the same assignment twice, small:

Using my RESEARCH ASSIGNMENT, give me only the five most important claims, each with one
source and the quoted supporting sentence. Mark each source primary or secondhand.
Stop at five.

Run it once in plain chat (no search), once with web search, and once in deep research / Research mode. Write down: how long it took, how many of the five quoted sentences you could find on the page, and how many sources were primary. That's your model-fit evidence. Chat without search will invent things. That's not a bug report, it's the reason the desk exists.

Card 6 · The gather run · slot: Act (first pass) → Ledger

Run this in deep research (ChatGPT) or with Research on (Claude):

Run my RESEARCH ASSIGNMENT under my SOURCE RULES.
Output a SOURCE LEDGER table with exactly the fields in my spec card, at least 15 rows.
Then list:
- claims you could not find a primary source for
- any two sources that disagree, with both sentences quoted
- whose voice is missing from the ledger
Do not write a conclusion yet.

Card 7 · The citation audit · slot: Check

You do this part. Open each URL yourself. For each row, mark one:

Then paste your marks back:

Here's my audit of your ledger: <paste rows with your marks>.
For every row marked wrong claim, dead or secondhand: find the primary source and the exact
supporting sentence, or delete the claim. Don't soften a claim to keep a row alive.
Then add this rule to the top of your answer and follow it in every future run:
"<the specific failure you caught, as a rule — e.g. 'never cite a proposed rule for what the
final rule says'>"

Copy that new rule into your Project's SOURCE RULES. The desk just learned something from its own failure.

Card 8 · Locate it in the primary · slot: Check

For the one claim your decision hangs on:

This claim is load-bearing: "<claim>". Find the primary document it rests on. Give me the
section, page or paragraph number and quote the sentence. Then tell me whether the primary
says MORE, LESS or SOMETHING DIFFERENT than the claim as written — be specific.

Card 9 · The independence collapse · slot: Check

For the finding "<finding>", list every source in the ledger that supports it. For each one,
trace the claim back to where it ORIGINATED: the press release, the filing, the wire story,
the dataset. Group sources that share an origin. Then tell me:
- how many links support it
- how many independent origins support it
- whether any origin is a party with a stake in the outcome

Card 10 · The contradiction table · slot: Structure → Check

Build a CONTRADICTION TABLE from the ledger. Fields fixed:
point of disagreement | source A + quoted sentence | source B + quoted sentence |
kind (different events / different definitions / different time / genuine conflict) |
what would resolve it | status (resolved / open)
Don't resolve anything by majority vote. Don't resolve anything by trusting the more
famous source. If it's open, say what evidence would close it.

Card 11 · The uncertainty map · slot: Structure

Sort every claim in the ledger into three bins:
KNOWN — a primary source says it, and nobody credible disputes it
CONTESTED — credible sources disagree; name them
UNKNOWN — nobody can know yet; name the event that would tell us, and its likely date if any
Then tell me which UNKNOWN matters most to the decision in my assignment, and why.

Card 12 · The critics' table · slot: Check (adversarial)

From ULM 2.2 LLM Expert Business Analysis (the bell hooks / Emma Goldman pass), pointed at a draft instead of a plan. Put these three in your Project instructions under CRITICS:

CRITICS
1. The skeptical domain expert: <a specific kind of expert for your field — e.g. an
   administrative-law professor who has read the whole rule>. Catches: misread mechanics,
   numbers compared that don't measure the same thing, definitions that slid.
2. Opposing counsel: argues the strongest honest case for the outcome my draft treats as
   less likely. Catches: assumptions that happen to match what I want.
3. The person most affected: <a specific person the decision lands on — e.g. someone who
   lives beside a freight corridor and drives a 2011 car>. Catches: whose costs aren't
   counted, whose voice the sources leave out, what the brief would sound like to them.
Each critic: raise at most 3 objections, each tied to a specific sentence in the draft and
to a source in the ledger (or flagged NO SOURCE). No general advice. No praise.

Then run:

Here's my draft brief: <paste>. Convene the CRITICS. Each critic gives their objections
separately. Then STOP. Don't revise the draft. I'll rule on each objection myself.

Card 13 · The critic diff · slot: Gate

You rule first, then hand it back:

My rulings: <objection 1: ACCEPT / REJECT + one-line reason> ... 
Revise the draft for the ACCEPTED objections only. Show me a diff: every sentence you
changed, before and after, and which objection it answers. Change nothing else.

Card 14 · Synthesis across the primaries · slot: Structure (the position)

From ULM 3.2 LLM Research Paper Synthesis — compress first, then synthesise, then fact-check the synthesis.

For each PRIMARY source in the ledger, write a 120-word digest that another model could use
to find where the sources agree and disagree: what it claims, on what basis, what it
explicitly does NOT claim, and its date. Then, using only those digests, state the
strongest position the evidence supports on my question, name where two primaries
disagree, and say which reading you took and why. Then fact-check your own position
against the ledger and mark any sentence with no ledger row as UNSUPPORTED.

Card 15 · The brief · slot: Act

Write the brief for <decision-maker>. One page. Structure fixed:
1. The decision and the recommendation in two sentences.
2. A table: claim | confidence (high / medium / low) | source (ledger row) — confidence is
   HIGH only if a primary source says it and it isn't contested.
3. What we don't know: the open items from the uncertainty map, each with what would tell us.
4. Whose voice is missing from this evidence.
5. What would change this recommendation.
No claim without a ledger row.

Card 16 · Run it twice: the diff · slot: Ledger

Run Card 6 again (a week later, or in the other product, or on the sibling question). Then:

Here are two SOURCE LEDGERS for the same assignment: RUN 1 <paste> and RUN 2 <paste>.
Tell me: claims in both · claims only in run 1 · claims only in run 2 · claims where the
two runs cite different sources or quote different sentences · any claim that changed
confidence. Don't merge them. I want the differences.

Card 17 · The watch list · slot: Rhythm

Save as a ChatGPT Task (weekly) or a Claude scheduled task if your plan has it, or paste it by hand on a calendar reminder.

WATCH: <assignment name>. Check only for NEW evidence since <last run date> on these items:
<paste your "what would change our mind" list>.
For each: CHANGED (quote the new sentence, link, date) or NO CHANGE.
Ignore new articles that repeat an origin already in the ledger — list them under
"same origin, not new."
If anything changed, say whether it moves the recommendation in the brief. Draft only.

Card 18 · The second desk · slot: all seven (transfer)

Write a fresh Spec Card and RESEARCH ASSIGNMENT for a different question in a different domain. Then, in a new Project:

Here's the spec card and assignment for a new research desk. Before you research anything,
compare it to this working desk's instructions: <paste desk 1's instructions>.
List what carries over unchanged, what has to change for the new domain (source rules,
critics, what counts as primary), and anything in the new assignment that's still vague.
Ask me about the vague parts. Then wait.

↑ Contents

Videos: Build Your AI Research Department

Block 1 fundamentals: Videos - Fundamentals. Times ±30 s.

Record these

1. A deep-research run, ChatGPT and Claude, then the citation audit on its output. No footage exists of any deep-research tool.
2. A cited brief made in the chat window (the existing one is MCP/LM Studio).

↑ Contents