zirui-song

Regenerate a Stata table and verify it in the compiled PDF

Refresh local .tex/.bib files against Overleaf before AI edits, so the agent never edits a stale copy (default Overleaf-to-Dropbox sync lags 10-20 min). Refresh path: probe `/project/<id>/updates` (read-only), skip if nothing changed, skip if all changes were our own Dropbox round-trips, else download-zip and extract only the web-origin changed files. WHEN: (1) before editing .tex/.bib/.cls/.sty/.bst under Apps/Overleaf/<project>/ run `overleaf-sync-now sync`; (2) on user request to refresh; (3) for first-time auth run `overleaf-sync-now login`. Claude Code: a PreToolUse hook auto-runs sync; manual invocation rarely needed. Codex CLI: invoke sync explicitly. AUTH RECOVERY when sync/setup fails: run `overleaf-sync-now login` (browser-assisted, works on Chrome 130+). Do NOT tell the user to 'log into Overleaf' in their daily browser — on Chrome 130+ app-bound encryption blocks on-disk cookie extraction regardless of login state. See body for full recovery flow.

Use to run a multi-round cross-model debate on an implementation plan — Claude drafts, Codex (gpt-5.4) critiques against the real codebase, Claude rebuts and synthesizes, optional second Codex pass — converging on a stronger plan. Combines codex-review-plan and review-codex-plan. Trigger phrases include "debate this plan with Codex", "go back and forth with Codex on the plan", "have Claude and Codex argue out the plan".

Use before interpreting any coefficient, writing up any result, or calling any null a finding — reports first-stage F, effective N, minimum detectable effect at 80% power, and the two most likely violations of the identifying assumption, so an underpowered design or weak first stage is stated up front rather than discovered by a referee. Trigger phrases include "is this null real", "am I underpowered", "MDE", "first stage", "before you interpret this".

Shared mechanics for calling the Codex CLI (gpt-5.4) to review plans. Reference for the codex-review-plan, review-codex-plan, and plan-debate skills — not invoked directly. Holds the wrapper invocation, reviewer prompt template, and gotchas.

Use before sending a design or draft to a coauthor or referee — launches five parallel adversarial subagents (identification, power, literature, data provenance, independent replication) against the research design in the repo, each citing file:line evidence, then synthesizes one threat memo ranked by probability-right × damage-if-right. Trigger phrases include "referee panel", "attack this design", "what would a referee say", "adversarial review of my paper".

Use after producing or updating empirical tables/figures that belong in a paper. Copies updated artifacts from the project's results/ directory to the corresponding Overleaf-synced folder, preserving filenames, and reports what changed so the LaTeX draft picks them up on next compile.

Use before the user reads any draft, memo, literature review, or results write-up — runs an adversarial verification pass that hunts factual errors, unsourced claims, and numbers that don't match the underlying code or output files. Every claim is either cited to an exact file:line or marked UNVERIFIED. Trigger phrases include "verify this draft", "check my numbers", "adversarial verification pass", "did you make this up".

Use before editing or running any Stata .do file in the user's research projects. Checks for missing `<your-stata-username>` username path branches, `cap`-suppressed esttab calls, and other recurring Stata pitfalls. Catches issues that otherwise surface only as silent failures (no log produced, lost estimates).

Use when running a multi-stage empirical pipeline (Stata + Python) to refresh tables/figures end-to-end. Performs pre-flight checks (Dropbox sync, column schemas, username path branches), executes scripts in order, verifies log files were produced, then summarizes key outputs and flags anomalies.

Use to map where a headline result lives and dies across the full space of defensible specifications — enumerates researcher degrees of freedom, runs a 200-500 spec grid in the background with checkpointing and resume, then produces a specification curve, a variance-decomposition table, a note on where the result flips, and a compiled online appendix. Trigger phrases include "specification curve", "overnight robustness sweep", "run every defensible spec", "how fragile is this result".

Use after drafting an implementation or research plan (in plan mode, or a ~/.claude/plans/ file) to get a second-opinion review from Codex (gpt-5.4) against the actual codebase, then critically incorporate the feedback. Makes the Claude Code plan better via cross-model review. Trigger phrases include "send the plan to Codex", "have Codex review this plan", "get a second opinion on the plan".

Use to have Claude critically review a Codex-authored implementation plan against the real codebase — or to have Codex draft a plan for a task first and then review it. The reverse of codex-review-plan. Trigger phrases include "review this Codex plan", "have Codex draft a plan then check it", "what's wrong with Codex's plan".

Use when preparing a paper for submission or responding to referees and you need a full battery of robustness checks for a headline result. Enumerates reviewer-anticipated variants across estimator, sample, controls, weights, and clustering, runs them as parallel jobs, and produces a comparison table + sign/significance stability summary.

Use to make every number in a paper machine-checkable — inventories scripts to the tables they produce, builds a YAML manifest of every numeric claim in the manuscript prose, writes verify_numbers.py to re-parse estimation logs and diff them against the text, adds make all / make verify targets and a pre-commit hook that blocks stale numbers. Trigger phrases include "self-verifying pipeline", "check every number in the paper", "which sentences went stale", "make verify".