Skills Reference

Skills are internal sub-agent skill files in skills/. They are not invoked directly by the user — commands load them as system prompts when spawning sub-agents. Each skill defines a narrow role, the files it reads, and the files it writes.

Pipeline skills

orchestrator.md

Read by: every slash command, before anything else.

Defines the execution rules that all agents must follow: how to run notebooks, the filesystem-as-state model, immutability rules, referee isolation protocol, and error handling.

Key rules enforced: - Agents communicate exclusively through files. No in-memory state passing. - data/paper_spec.json is read-only after stage 1. - code_run/ is written by nbconvert only. - paper/review_history/ is append-only.

notebook_runner.md

Read by: /run, /recast, /recast-cf, /stage.

Executes a single notebook:

jupyter nbconvert --to notebook --execute --inplace \
  --ExecutePreprocessor.timeout=1800 \
  --output-dir <project>/code_run/ \
  <project>/code_build/0N_*.ipynb

Notebooks must be executed with cwd = <project>/code_build/ because they import from paths import * to resolve all file paths. On failure, reports the failing cell number and error traceback.

Gate

advisor_gate.md

Read by: /recast, /recast-cf (after stage 6).

Spawns three independent checks in sequence. Each check reads specific files and returns PASS or FAIL with a reason.

Check Files read Validates
Code Auditor All code_run/ notebooks No execution errors; all expected outputs exist
Identification Checker paper_spec.json, replication_check.json Internally consistent identification strategy
Data Validator diagnostics_flags.json No critical data quality flags

All three checks must return PASS. The orchestrator stops the pipeline on the first FAIL.

Review skills

review_loop.md

Read by: /review.

Defines the loop logic: state detection (count existing round_*/ dirs), per-round execution (3 referees → synthesis → revision), exit conditions, and progress reporting format.

revision_agent.md

The only agent that writes to paper/paper.tex and code_build/ notebooks.

Reads the synthesis report and implements changes:

Issue severity Action
Blocking Edits code_build/0N_*.ipynb (appends ## Revision Round N section) + signals RERUN_NEEDED: yes
Major Direct edit to paper/paper.tex
Minor Direct edit to paper/paper.tex

After implementing all changes, writes paper/review_history/round_N/changelog_N.md documenting every change made and every issue deferred, with reasons.

Key rule: never overwrites prior cells in notebooks — always appends. Never overwrites prior files in review_history/.

synthesis_referee.md

Read by: review loop, after all three referee reports are written.

The first agent in the chain that sees all three referee reports. Deduplicates overlapping concerns, classifies each unique issue as blocking / major / minor, and checks prior changelog_*.md files to avoid re-raising resolved issues.

final_referee.md

Read by: /final.

Reads all rounds’ ref*.md, synthesis.md, and changelog_*.md files, plus the final paper/paper.tex. Writes paper/review_history/final_report.md as a human-readable summary.

Referees

Each referee is isolated: it receives only its own skill file as a system prompt and reads only the files listed below. No referee has access to any other referee’s output. The synthesis referee is the first agent that sees all three reports.

referee_1_identification.md

Mandate: Causal identification strategy only.

Reads: paper_spec.json, data/results/replication_results.json, paper/paper.tex

Evaluates: instrument validity, exclusion restriction, first-stage strength, LATE vs. ATE interpretation, external validity.

referee_2_dml_methods.md

Mandate: DoubleML / Causal Forest implementation and methodological choices.

Reads: paper_spec.json, data/results/dml_results.json and/or causal_forest_results.json, hte_results.json, code_build/04_*.ipynb, paper/paper.tex

Evaluates: cross-fitting procedure, learner selection, nuisance R², GATE/CATE validity. For Causal Forest: honesty, tree count, SE plausibility (check 21 — blocking if ATE CI is 10x+ narrower than individual CATE CIs), feature importance interpretation.

referee_3_robustness.md

Mandate: Replication fidelity and robustness.

Reads: paper_spec.json, data/results/replication_check.json, data/results/diagnostics_flags.json, paper/paper.tex

Evaluates: numerical match to published tables, sample coverage, sensitivity to specification, data quality flags from stage 5.