brycewang-stanford/aer-workflow

AER Workflow

Overview

This is the router. It does not replace any specialized skill. It tells you which one to use next, and in what order.

Default assumption: unless the user names a different venue, the manuscript targets AER, AER: Insights, or an AEJ journal — not a finance journal, not a generic economics field journal, and not a working-paper repository.

When to Use

• The user asks "what should I work on next?"
• The user dumps a draft and you must decide where the bottleneck is
• The user asks for a full paper to be drafted end-to-end and you need the gate sequence
• The user is rotating between writing, empirics, and revision and loses track of which stage they are in
• A new reviewer report has arrived and the work mode must switch from drafting to rebuttal

Routing Map

Use:

• aer-topic-selection when the project is new, when the user is undecided between AER / Insights / an AEJ, or when the contribution sentence cannot be written in one line
• aer-literature when the closest papers must be mapped (novelty scan, antecedents paragraph) or when any citation needs verifying — for AI-drafted manuscripts, every citation
• aer-identification when the empirical design is the bottleneck — DiD, IV, RDD, SCM, shift-share, event study, RCT analysis
• aer-preregistration when the study runs a field/lab/survey experiment or collects primary data — write the pre-analysis plan, size the sample from a power/MDE calculation, and register with the AEA RCT Registry before the intervention
• aer-robustness when the main results exist but referee-anticipating checks (placebo, heterogeneity, mechanism, alternative samples) are missing or weak
• aer-paper-body when drafting or repairing the body sections — background, data, empirical strategy, results narration, magnitude interpretation, mechanisms, conclusion
• aer-introduction when drafting or rewriting the introduction, or when the abstract is over 100 words
• aer-tables-figures when regression tables are inconsistent, oversized, footnote-bloated, or do not match AER house style
• aer-consistency when the assembled manuscript needs the integrity audit — numbers vs. tables, sample sizes, conversions, cross-references, citation two-way match
• aer-referee-sim when a complete draft needs the adversarial desk screen and three simulated referee reports before submission
• aer-replication when preparing the AEA Data and Code Availability deposit, writing the README, or auditing reproducibility before acceptance
• aer-submission when running the final preflight before clicking submit — length, format, cover letter, conflicts
• aer-rebuttal when reviewer comments exist and a point-by-point response letter plus aligned manuscript edits are needed
• aer-statspai (optional engine) when the design is already fixed and the user wants to run the empirics from the agent — a unified Python API / MCP server covering DiD, IV, RDD, SCM, robustness, and table export. It executes what aer-identification chose; it does not replace that choice.

Default Sequence

For most empirical AER-track manuscripts, prefer this order:

1. aer-topic-selection — fix the contribution sentence and the target venue before anything else
2. aer-literature — map the 3-6 closest papers; kill the project now if it is scooped
3. aer-identification — stress-test the design; if it fails here, no later skill saves the paper
   • Experimental / primary-data branch: if the design randomizes treatment or collects new data, detour to aer-preregistration before the intervention — pre-analysis plan, power/MDE-justified sample size, AEA RCT Registry timestamp. Skip this branch for observational data already realized.
4. aer-robustness — anticipate the three robustness checks the median referee will demand (run the empirics via templates/ or aer-statspai)
5. aer-paper-body — draft background, data, strategy, results, mechanisms, conclusion
6. aer-introduction — only now write the five-paragraph intro and the 100-word abstract, summarizing a body that exists
7. aer-tables-figures — finalize the main exhibits in AER house style
8. aer-consistency — the integrity audit; every number, reference, and citation must reconcile
9. aer-referee-sim — adversarial internal review; loop back to the routed fix skill until the verdict is at least major R&R on a fresh run
10. aer-replication — assemble the deposit package while results are still fresh in code
11. aer-submission — final preflight: length, format, cover letter, COI
12. aer-rebuttal — after external review, revise manuscript first, then write the letter against the revised version

Quality Gates for End-to-End Drafting

When the user asks the agent to write the paper, the sequence above is a set of gates, not suggestions. Do not advance past a gate that fails:

• Gate A (after step 3): contribution sentence written, venue chosen, design survives aer-identification red flags
• Gate B (after step 5): every claim in the draft body traces to an exhibit or a verified citation
• Gate C (after step 8): aer-consistency reports all-pass; aer-literature citation ledger fully closed
• Gate D (after step 9): aer-referee-sim verdict ≥ major R&R on two consecutive fresh runs
• Gate E (after step 11): aer-submission preflight all green

Decision Cues

If the user says...

• "I have an idea but I don't know if it's AER-worthy" → aer-topic-selection
• "Has someone already done this?" / "Are these references real?" → aer-literature
• "My DiD has staggered treatment" → aer-identification
• "My first stage F is 8" → aer-identification (weak-IV branch)
• "I'm about to run a field experiment" / "What sample size do I need?" / "Where do I register the RCT?" → aer-preregistration
• "The referee said the result might be driven by X" → aer-robustness
• "My results section just walks through the tables" → aer-paper-body
• "What does a 0.042 coefficient on log wages mean?" → aer-paper-body (magnitude rules)
• "My intro is 4 pages and the editor hated it" → aer-introduction
• "My table 3 has 14 columns" → aer-tables-figures
• "Do the numbers in my abstract match my tables?" → aer-consistency
• "Tear this draft apart before I submit" → aer-referee-sim
• "The Data Editor flagged my README" → aer-replication
• "I'm submitting tomorrow" → aer-submission
• "I got an R&R with three reports" → aer-rebuttal
• "Just run my staggered DiD / IV / RDD for me" → aer-statspai (after the design is settled in aer-identification)

Common Mistakes

• polishing the introduction before the identification strategy is stable
• drafting the body before the empirics exist, then bending results prose to fit promises
• writing the abstract before deciding AER vs AER:Insights (the 100-word limit is the same but the framing differs sharply)
• treating tables and figures as a final-week task; AER tables drive how reviewers read results
• submitting without the consistency audit — text-table mismatches are the cheapest rejection trigger there is
• skipping the referee simulation because the deadline is close; the real referees are slower and harsher
• writing the rebuttal letter against the old draft instead of the revised manuscript
• assembling the replication package only after acceptance — the AEA Data Editor's report is now part of the production timeline and can delay publication by weeks
• defaulting to the OLS + cluster-by-state recipe; modern AER demands design-based identification

Anti-Patterns

• Using a generic "scientific writing" skill in place of aer-introduction or aer-paper-body — the conventions (no intro heading, five-paragraph Head formula, finding-first results paragraphs, magnitude discipline) are economics-specific
• Using a generic "data availability" skill in place of aer-replication — the AEA policy is unusually strict and the openICPSR workflow is specific
• Skipping aer-topic-selection because "I already know this is an AER paper" — desk rejection across the AER family is high, and the top-5 bar is cross-subfield interest, not technical competence
• Letting any skill write citations from memory — aer-literature's verification protocol is mandatory for machine-drafted text
• Treating aer-referee-sim praise as evidence of quality; its only export is the revise list

Repository Resources

Bundled with the installed skill, no repository checkout needed --- read it before the repo resources below:

• references/gate-map.md --- the 12-step lifecycle with entry criteria, gates, and artifacts per step

When working from the AER-skills repository or plugin bundle, load only the relevant resource:

• Full workflow diagram: docs/workflow-map.md
• Editor/referee no-go checklist: docs/desk-rejection-audit.md
• Estimator and diagnostic defaults: docs/methods-reference.md
• Prose rules shared by the writing skills: docs/style-guide.md
• Scoring rubric for the internal review gate: docs/referee-report-rubric.md
• Policy source register: docs/source-register.md
• SkillOpt-style improvement protocol: docs/skillopt-evaluation-protocol.md
• Held-out routing scenarios: examples/skillopt-routing-scenarios.json

Skill Optimization Loop

When improving this skill stack, treat aer-workflow as the router under test. Use docs/skillopt-evaluation-protocol.md to collect failures, keep each patch bounded, and accept routing or gate changes only after python3 scripts/run_skillopt_gate.py and make preflight pass. Do not tune the router by naming only the expected skill; use user-like manuscript prompts from examples/skillopt-routing-scenarios.json or add a new scenario first.

Handoff Contract

Whenever this skill is invoked, end with:

NEXT SKILL: <aer-skill-name>
REASON: <one sentence>
INPUTS NEEDED: <list of artifacts the next skill needs>
GATE STATUS: <which quality gates are passed / open>

This keeps the agent loop tight when the user runs multiple skills in sequence.