Slash Commands

All commands are defined in .claude/commands/ and are invoked from Claude Code. Every command reads skills/orchestrator.md first before taking any action.

Command What it does
/recast Full pipeline with DoubleML extension
/recast-cf Full pipeline with Causal Forest extension
/run Same as /recast (legacy alias)
/init Scaffold a new project folder
/stage Re-run from stage N onwards
/review Review loop only
/final Final referee report only
/publish Publish a finished paper to the website

/recast

/recast <project_path>

Full pipeline with DoubleML extension: stages 1–6 → Advisor Gate → Review Loop → Final Report. Uses 04_dml_extension.ipynb at stage 4 (PLIV/PLR with RandomForest + LassoCV learners).

/recast-cf

/recast-cf <project_path>

Full pipeline with Causal Forest extension: stages 1–3, then 04_causal_forest.ipynb at stage 4, then stages 5–6 → Advisor Gate → Review Loop → Final Report. Uses CausalForestDML (OLS/DID) or CausalIVForest (IV with binary instrument) from EconML.

The causal forest pipeline does not produce dml_results.json. Diagnostics and the report auto-detect which results exist and adapt accordingly.

/run

/run <project_path>

Same as /recast — the legacy command name. Full pipeline with DoubleML.

Steps:

  1. Validates that <project_path> contains .dml_project and config.yaml. If not, prompts you to run /init first.
  2. Reads config.yaml and skills/orchestrator.md.
  3. Runs stages 1–6 sequentially via notebook_runner. Stops immediately on any stage failure and reports the failing cell and error.
  4. Runs the Advisor Gate (three independent checks). Stops if any check fails.
  5. Runs the Review Loop (up to config.yaml > review.max_rounds rounds).
  6. Runs the Final Referee.
  7. Prints a summary: stages completed, rounds of review, final verdict, paths to key outputs.

Key rules: - The Advisor Gate is never skipped. - The Final Referee always runs, even if the review loop exits cleanly after round 1. - code_run/ receives executed notebook output — never edited directly.

/init

/init <project_path>

Scaffolds a new project folder at <project_path>.

Creates:

<project_path>/
├── .dml_project
├── config.yaml              ← from config_template.yaml
├── raw_data/
├── data/results/
├── code_build/              ← notebooks + paths.py copied from framework
├── code_run/
└── paper/
    ├── tables/ figures/
    ├── prompts/             ← referee skill files copied here
    └── review_history/

If <project_path> already exists and contains .dml_project, aborts with a warning.

/stage

/stage <N> <project_path>

Re-runs the pipeline from stage N through stage 6, then runs the Advisor Gate and Review Loop.

N must be between 1 and 6. Use this after fixing a specific issue:

Scenario Command
Fixed the data cleaning /stage 2 <project>
Changed a DML specification /stage 4 <project>
Revised paper.tex manually /stage 6 <project>
Want to re-run only the review /review <project>

Note: stages before N are not re-run. Their output files must already exist in data/.

/review

/review <project_path>

Runs the Review Loop only. Assumes stages 1–6 and the Advisor Gate have already passed.

Use this when you have manually edited paper/paper.tex and want to put it through the peer review process without re-running the analysis.

Reads skills/orchestrator.md and skills/review_loop.md. State is determined by counting paper/review_history/round_*/ directories — it resumes from the next round if prior rounds already exist.

/final

/final <project_path>

Runs the Final Referee only. Reads all prior rounds’ referee reports, syntheses, and changelogs, and writes paper/review_history/final_report.md.

Use this to regenerate the final report without re-running the review loop, for example after manually editing the changelog.

/publish

/publish <project_path>

Generates a complete, populated website page for a finished pipeline project and copies all required assets into website/papers/<slug>/.

Requires the pipeline to be fully complete: final_report.md must exist (run /final <project_path> first if needed).

Steps:

  1. Reads data/paper_spec.json, data/results/dml_results.json, data/results/replication_check.json, data/results/replication_results.json, and paper/review_history/final_report.md to extract all metadata.
  2. Derives the slug (lowercasesurname + year, e.g. acemoglu2001).
  3. Creates website/papers/<slug>/ and writes a fully populated index.qmd with correct YAML frontmatter (title, authors, DML estimates, replication status, categories, etc.) and populated replication and DML tables.
  4. Copies forest_plot.png and all five referee report .md files into place.

Output:

website/papers/<slug>/
├── index.qmd           ← populated page, ready for Quarto
├── forest_plot.png
└── reports/
    ├── ref1.md
    ├── ref2.md
    ├── ref3.md
    ├── synthesis.md
    └── final_report.md

Push to main after running — GitHub Actions rebuilds the site automatically.

Key rules:

  • Aborts if website/papers/<slug>/ already exists (pass --force to overwrite).
  • Never modifies referee .md files — copies them verbatim.
  • If any JSON field is missing, writes ??? as a placeholder and adds an HTML comment <!-- TODO: fill in --> so you can find it.
  • Does not copy paper.pdf (may be large); reminds you to add it manually if you want a download link.