End-to-End Research with AI

Labour-market-AI-exposure case — question to paper, one checkpoint at a time.

New to this? Start here.

No coding background needed. Here's what a few names you'll see mean, in plain language, with a link to where you'd actually get each one.

The whole two-day pipeline as one graph, plus the reasoning rule inside every step — the same decision logic this app runs on the built-in labour-market-AI-exposure case. This page documents the pedagogy; the pipeline above starts once you begin.

Tool presentations

A 12-slide deck per tool — what it is, which stage it's used in, and how it fits with the rest.

Decision graphs on this page

Jump straight to a step's reasoning graph below.

flowchart TD
    subgraph Day1["Day 1 - Tool fluency"]
        D1_1["1. Environment check"]
        D1_2["2. Write your AGENTS.md"]
        D1_3["3. Literature & verification tools"]
        D1_4["4. Build a referee-style skill"]
        D1_5["5. Six-subagent pattern"]
        D1_1 --> D1_2 --> D1_3 --> D1_4 --> D1_5
    end

    subgraph Day2["Day 2 - Full pipeline"]
        D2_1["1. Hypothesis"]
        D2_2["2. Design & identification"]
        D2_3["3. Data & analysis lab"]
        D2_4["4. Results & robustness"]
        D2_5["5. Manuscript (voice.md)"]
        D2_6["6. Governance & disclosure"]
        D2_1 -->|one-sentence hypothesis| D2_2
        D2_2 -->|design memo + flagged alternative| D2_3
        D2_3 -->|naive vs. robust estimate| D2_4
        D2_4 -->|numbers & figures| D2_5
        D2_5 -->|draft| D2_6
        D2_6 -.->|referee or disclosure gate fails| D2_4
    end

    D1_5 ==>|Day 1 complete| D2_1
    D1_2 -.->|governs every step| Day2
    D1_5 -.->|identification-skeptic| D2_1
    D1_5 -.->|identification-skeptic| D2_2
    D1_5 -.->|identification-skeptic, stats-reviewer| D2_3
    D1_5 -.->|writing-editor| D2_5
    D1_5 -.->|disclosure-auditor| D2_6
    D1_3 -.->|citation-verification habit| D2_5
    D1_4 -.->|same skill, applied to the real draft| D2_6
  

Day 1 — Tool fluency

Five steps: get fluent with the mechanics before turning them on your own research question.

flowchart TD
    A(["Run python scripts/check_setup.py"]):::start --> B{"All checks pass?"}:::decision
    B -->|No| C["Fix by category:
packages missing -> pip install -r requirements.txt
claude not found -> npm install -g @anthropic-ai/claude-code
no key -> export ANTHROPIC_API_KEY"]:::no
    C --> A
    B -->|Yes| D["Ask the agent: 'Read AGENTS.md and summarize
the ground rules in 3 bullets.'"]:::yes
    D --> E{"Does the summary match the real file?"}:::decision
    E -->|No| F["Agent isn't picking up repo instructions -
fix config before relying on it"]:::no
    E -->|Yes| G(["Ready for the rest of the course"]):::final

    classDef start fill:#eef2f6,stroke:#16324f,color:#16324f;
    classDef decision fill:#ffffff,stroke:#c8762a,color:#0e2336;
    classDef yes fill:#e3f1e8,stroke:#2f7a4f,color:#2f7a4f;
    classDef no fill:#f8e4e2,stroke:#a13a3a,color:#a13a3a;
    classDef final fill:#16324f,stroke:#16324f,color:#ffffff;
  

flowchart TD
    subgraph SG1["Precedence: which file wins"]
    A1(["About to ask the agent to do a task on your angle"]):::start --> B1{"Does your angle's AGENTS.md
say anything about this?"}:::decision
    B1 -->|Yes| C1["Follow the angle-specific rule - more specific wins"]:::yes
    B1 -->|No| D1["Fall back to root AGENTS.md:
no invented citations, no bare causal claims,
disclose every AI step"]:::no
    end

    subgraph SG2["Test it"]
    A2(["Give the agent that same task"]):::start --> B2{"Did it visibly follow your added rule,
not just the root one?"}:::decision
    B2 -->|No| C2["Rule was too vague - tighten wording, retry"]:::no
    B2 -->|Yes| D2(["Done - the file changes behavior observably"]):::final
    end

    classDef start fill:#eef2f6,stroke:#16324f,color:#16324f;
    classDef decision fill:#ffffff,stroke:#c8762a,color:#0e2336;
    classDef yes fill:#e3f1e8,stroke:#2f7a4f,color:#2f7a4f;
    classDef no fill:#f8e4e2,stroke:#a13a3a,color:#a13a3a;
    classDef final fill:#16324f,stroke:#16324f,color:#ffffff;
  

flowchart TD
    A(["Agent is about to cite a paper"]):::start --> B{"Does it come from an actual search result?
(script output, or Consensus/Elicit/Scite you pasted in)"}:::decision
    B -->|No| C["Say 'unverified' explicitly -
never soften into a plausible-sounding citation"]:::no
    B -->|Yes| D{"In Scite: has it been contradicted
by later work?"}:::decision
    D -->|Yes| E["Report both the paper and the
contradicting evidence - don't cherry-pick"]:::no
    D -->|No| F{"Does the Consensus/Elicit synthesis
match the paper's own abstract?"}:::decision
    F -->|No| G["Flag the mismatch - trust the
primary source over the AI summary"]:::no
    F -->|Yes| H(["Safe to cite - note venue and year"]):::final

    classDef start fill:#eef2f6,stroke:#16324f,color:#16324f;
    classDef decision fill:#ffffff,stroke:#c8762a,color:#0e2336;
    classDef yes fill:#e3f1e8,stroke:#2f7a4f,color:#2f7a4f;
    classDef no fill:#f8e4e2,stroke:#a13a3a,color:#a13a3a;
    classDef final fill:#16324f,stroke:#16324f,color:#ffffff;
  

flowchart TD
    A(["Run /referee-review on a manuscript draft"]):::start --> B{"1. Causal language present -
does the paragraph name an identification strategy?"}:::decision
    B -->|No| B2["Flag: soften to correlational language,
or add a forward reference"]:::no
    B -->|Yes| C{"2. Any post x exposure estimate -
is there pre-trend evidence?"}:::decision
    C -->|No| C2["Flag - the single most important
thing this skill exists to catch"]:::no
    C -->|Yes| D{"3. Every citation traceable
to a real search result?"}:::decision
    D -->|No| D2["Flag as unverified, per AGENTS.md"]:::no
    D -->|Yes| E{"4. Do 2-3 spot-checked numbers match
the analysis script output?"}:::decision
    E -->|No| E2["Hard blocker, not a style note"]:::no
    E -->|Yes| F{"5. Does every AI-assisted step have
a disclosure entry?"}:::decision
    F -->|No| F2["Flag"]:::no
    F -->|Yes| G(["Ready for governance sign-off: yes
(no, if 1-3 above unresolved)"]):::final

    classDef start fill:#eef2f6,stroke:#16324f,color:#16324f;
    classDef decision fill:#ffffff,stroke:#c8762a,color:#0e2336;
    classDef yes fill:#e3f1e8,stroke:#2f7a4f,color:#2f7a4f;
    classDef no fill:#f8e4e2,stroke:#a13a3a,color:#a13a3a;
    classDef final fill:#16324f,stroke:#16324f,color:#ffffff;
  

flowchart TD
    A(["You need a specific kind of check on your work"]):::start --> B{"What are you worried might be wrong?"}:::decision
    B -->|Citation not verified| C1["literature-scout"]:::yes
    B -->|Causal claim might be a confound| C2["identification-skeptic"]:::yes
    B -->|Data code might have a silent bug| C3["data-integrity-checker"]:::yes
    B -->|Statistics might not match the question| C4["stats-reviewer"]:::yes
    B -->|Prose needs tightening| C5["writing-editor"]:::yes
    B -->|An AI step might be undisclosed| C6["disclosure-auditor"]:::yes
    C1 --> D{"Run 2+ relevant subagents independently -
do they flag the same thing?"}:::decision
    C2 --> D
    C3 --> D
    C4 --> D
    C5 --> D
    C6 --> D
    D -->|Yes, rare| E1["Probably an obvious, high-priority issue"]:::no
    D -->|No, expected| E2(["That's the point - independent blind spots"]):::final

    classDef start fill:#eef2f6,stroke:#16324f,color:#16324f;
    classDef decision fill:#ffffff,stroke:#c8762a,color:#0e2336;
    classDef yes fill:#e3f1e8,stroke:#2f7a4f,color:#2f7a4f;
    classDef no fill:#f8e4e2,stroke:#a13a3a,color:#a13a3a;
    classDef final fill:#16324f,stroke:#16324f,color:#ffffff;
  

Day 2 — Full pipeline

Six steps, hypothesis to governance — this is what the pipeline above actually runs once you begin.

flowchart TD
    A(["Generate 3 distinct hypotheses"]):::start --> B{"Is there a result that would
prove this one wrong?"}:::decision
    B -->|No| C["Not a real hypothesis - reject or reword"]:::no
    B -->|Yes| D{"identification-skeptic:
is there a confound producing the same
pattern with no true AI effect?"}:::decision
    D -->|Yes| E["Name the alternative explanation -
narrow or weaken the hypothesis"]:::no
    D -->|No| F["Refine: name outcome, comparison,
time window, direction"]:::yes
    E --> G{"Take it through the confound
question once more - survives twice?"}:::decision
    F --> G
    G -->|No| H["Refine again"]:::no
    G -->|Yes| I(["In good enough shape to move to design"]):::final

    classDef start fill:#eef2f6,stroke:#16324f,color:#16324f;
    classDef decision fill:#ffffff,stroke:#c8762a,color:#0e2336;
    classDef yes fill:#e3f1e8,stroke:#2f7a4f,color:#2f7a4f;
    classDef no fill:#f8e4e2,stroke:#a13a3a,color:#a13a3a;
    classDef final fill:#16324f,stroke:#16324f,color:#ffffff;
  

flowchart TD
    A(["Write the design memo: comparison, counterfactual,
estimating equation, what would violate the assumption"]):::start --> B{"Does it name a violation specific
to THIS dataset?"}:::decision
    B -->|No| C["Too generic - name a data-specific risk
(hint: check pre-trends)"]:::no
    B -->|Yes| D["Run the memo past identification-skeptic"]:::yes
    D --> E{"Does it flag a specific vulnerability?
(e.g. post x exposure, no pre-trend check)"}:::decision
    E -->|Yes| F(["It proposes an alternative spec -
test it in the data lab"]):::final
    E -->|No| G(["Proceed to the data lab as specified -
document what was checked"]):::final

    classDef start fill:#eef2f6,stroke:#16324f,color:#16324f;
    classDef decision fill:#ffffff,stroke:#c8762a,color:#0e2336;
    classDef yes fill:#e3f1e8,stroke:#2f7a4f,color:#2f7a4f;
    classDef no fill:#f8e4e2,stroke:#a13a3a,color:#a13a3a;
    classDef final fill:#16324f,stroke:#16324f,color:#ffffff;
  

flowchart TD
    A(["Run the naive ai_exposure x post2023 estimate;
plot pre-2023 hours by exposure tercile"]):::start --> B{"Are the pre-2023 lines parallel
across terciles?"}:::decision
    B -->|No| C["Confound present - add a continuous
trend control (ai_exposure x year)"]:::no
    B -->|Yes| C2["Naive estimate more trustworthy -
still disclose the check"]:::yes
    C --> D{"Once the control is added,
does the estimate move a lot?"}:::decision
    D -->|Yes| E(["That gap is the main result -
write it up, don't bury it"]):::final
    D -->|No| F(["Report both numbers - the check
itself is the evidence"]):::final
    E --> G{"Send the spec to identification-skeptic
and stats-reviewer - anything new?"}:::decision
    F --> G
    G -->|Yes| H["Address it before write-up"]:::no
    G -->|No| I(["Proceed to results"]):::final

    classDef start fill:#eef2f6,stroke:#16324f,color:#16324f;
    classDef decision fill:#ffffff,stroke:#c8762a,color:#0e2336;
    classDef yes fill:#e3f1e8,stroke:#2f7a4f,color:#2f7a4f;
    classDef no fill:#f8e4e2,stroke:#a13a3a,color:#a13a3a;
    classDef final fill:#16324f,stroke:#16324f,color:#ffffff;
  

flowchart TD
    A(["Report the naive estimate, the pre-trend evidence,
and the robust estimate with the added control"]):::start --> B{"How far did the estimate move?"}:::decision
    B -->|A lot| C(["Foreground the gap as the main finding -
state confidence caveats"]):::final
    B -->|Little| D(["Note robustness confirms the initial read"]):::final
    C --> E{"Optional checks (median split, drop covid,
occupation trend) - estimate changes materially?"}:::decision
    D --> E
    E -->|Yes| F["Report it - don't drop the check
that ruins the nicer number"]:::no
    E -->|No| G["Note stability across specifications"]:::yes

    classDef start fill:#eef2f6,stroke:#16324f,color:#16324f;
    classDef decision fill:#ffffff,stroke:#c8762a,color:#0e2336;
    classDef yes fill:#e3f1e8,stroke:#2f7a4f,color:#2f7a4f;
    classDef no fill:#f8e4e2,stroke:#a13a3a,color:#a13a3a;
    classDef final fill:#16324f,stroke:#16324f,color:#ffffff;
  

flowchart TD
    A(["Before editing anything"]):::start --> B{"Does voice.md have concrete
examples, not generic advice?"}:::decision
    B -->|No| C["Fill it in with real, specific examples first"]:::no
    B -->|Yes| D["writing-editor reads it, then matches it"]:::yes
    D --> E{"Does the edit touch causal-confidence wording,
or add a citation/number/claim not already there?"}:::decision
    E -->|Yes| F["Never - that's substance,
route to identification-skeptic + author"]:::no
    E -->|No| G{"Is the sentence ambiguous about
correlation vs. causation?"}:::decision
    G -->|Yes| H["Flag it - don't silently pick a reading"]:::no
    G -->|No| I{"Is this a style edit or a substance edit?"}:::decision
    I -->|Unsure| J["Treat as substance - ask the author, don't guess"]:::no
    I -->|Style, e.g. hedging/throat-clearing| K{"Does voice.md call for that register?"}:::decision
    K -->|Yes| L["Keep it as-is"]:::yes
    K -->|No| M["Cut it"]:::yes
    M --> N(["Proceed - preserve paragraph structure"]):::final
    L --> N

    classDef start fill:#eef2f6,stroke:#16324f,color:#16324f;
    classDef decision fill:#ffffff,stroke:#c8762a,color:#0e2336;
    classDef yes fill:#e3f1e8,stroke:#2f7a4f,color:#2f7a4f;
    classDef no fill:#f8e4e2,stroke:#a13a3a,color:#a13a3a;
    classDef final fill:#16324f,stroke:#16324f,color:#ffffff;
  

flowchart TD
    A(["disclosure-auditor scans the repo for
evidence of AI-assisted steps"]):::start --> B{"Is there a matching entry in
this checklist for each one?"}:::decision
    B -->|No| C["Flag the gap - add the entry now
(asked / tool / human check),
not reconstructed from memory later"]:::no
    B -->|Yes, but vague| D["Flag as insufficient -
'used AI for writing' isn't an entry"]:::no
    B -->|Yes, specific| E(["Entry passes"]):::final
    A --> F{"Is there a step that plainly
wasn't AI-assisted?"}:::decision
    F -->|Yes| G["No entry needed just because AI
is used elsewhere in the repo"]:::yes

    classDef start fill:#eef2f6,stroke:#16324f,color:#16324f;
    classDef decision fill:#ffffff,stroke:#c8762a,color:#0e2336;
    classDef yes fill:#e3f1e8,stroke:#2f7a4f,color:#2f7a4f;
    classDef no fill:#f8e4e2,stroke:#a13a3a,color:#a13a3a;
    classDef final fill:#16324f,stroke:#16324f,color:#ffffff;
  

This app streamlines the 11-step teaching sequence above into a 9-stage pipeline (dataset → literature → hypothesis → design → analysis → results → manuscript → referee → disclosure) — see the hosted-app README's "What's simplified" section for exactly what's folded together.