review-doc-commit
Review, document, and commit. Use proactively when user asks to commit. Hard gate: no commit until review is clean.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
メニュー
Review, document, and commit. Use proactively when user asks to commit. Hard gate: no commit until review is clean.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Guide for rigorous economic data analysis. Use PROACTIVELY whenever performing data analysis on economic or financial datasets — importing, cleaning, merging, constructing variables, or producing summary statistics. Three core principles: (1) describe before and after every transformation, (2) document in jupytext percent format with interleaved code/narrative/outputs, (3) validate against economic intuition, literature, and cross-variable relationships. Includes pitfall checklists for merges, aggregations, filtering, and variable construction. Language-agnostic (Python, Julia). Trigger: any data analysis task involving economic, financial, or panel data.
Search and download financial data from WRDS (Wharton Research Data Services). Use when asked to "download from WRDS", "search WRDS", "get CRSP data", "download Compustat", "find WRDS table", "get stock returns", "download IBES", or any WRDS/financial database task involving CRSP, Compustat, IBES, TAQ, OptionMetrics, Fama-French, BoardEx, DealScan, or other WRDS datasets.
Comprehensive academic paper review covering mathematical correctness, writing clarity, consistency, argumentation, proofreading, and citations. Use when user asks to 'review draft', 'check paper', 'proofread manuscript', or requests feedback on academic writing. Can also verify code-paper consistency when source code is available. Defaults to comprehensive + standard review, with optional deep parallel review when Codex multi-agent support is available.
Convert PDFs to Markdown using Mistral OCR API with image extraction. Use when you need to extract structured text and images from PDFs, especially for scanned documents or documents with complex formatting. Outputs Markdown with embedded images.
Create new academic research projects with two-folder architecture. Use when user wants to create a new research project, start a new paper, set up a new analysis project, or mentions needing a project structure for research.
Create formal, fact-checked work journal entries after completing analysis work. Use when user asks to "summarize work", "document results", or "create work journal entry". Ensures code is committed, copies figures to attachments, and creates objective summaries with mandatory citations plus a report-quality verification pass. For quick reports without fact-checking, use the `report-in-markdown` skill.
| name | review-doc-commit |
| user-invocable | true |
| description | Review, document, and commit. Use proactively when user asks to commit. Hard gate: no commit until review is clean. |
Review, document, and commit workflow. Use dedicated subagents for review, documentation, and commit phases.
Execution model:
Parse the user's request:
/review-doc-commit with no path → all unstaged/staged changes/review-doc-commit <message or path> → scope to the specified path or use as contextRun git status and git diff (staged + unstaged) to identify all changed files. If a path scope was given, filter to only changes under that path.
Output contract from Scope phase:
Review and update documentation thoroughly so it reflects the latest changes. This includes inline docs and top-level/project docs.
Coverage checklist:
CLAUDE.md / AGENTS.md (see nested structure below)README.md (root and impacted module-level READMEs)docs/, architecture notes, runbooks, examples, changelogs)Context is progressively revealed through a hierarchy of CLAUDE.md files. Each level adds module-specific guidance without repeating what parent docs already cover:
CLAUDE.md — overall architecture, tech stack, build/test commands, project-wide conventionsCLAUDE.md — the module's purpose, its conventions, non-obvious design decisions, and how to work with itWhen reviewing or documenting a file, walk up from the file's directory to the repo root and read every CLAUDE.md encountered along the way. The union of these files provides the full context for that file.
AGENTS.md is a mirror of CLAUDE.md. Both names should resolve to the same content. When creating or discovering guidance docs:
CLAUDE.md exists, create a symlink: ln -s CLAUDE.md AGENTS.mdAGENTS.md exists, create a symlink: ln -s AGENTS.md CLAUDE.mdCLAUDE.md exists.CLAUDE.md, create one describing the module's purpose and conventions.AGENTS.md symlink exists alongside every CLAUDE.md (and vice versa).Rules for documentation content:
Output contract from Documentation Subagent:
Phase 2 must complete before Phase 3 starts. This ensures reviewers can verify that documentation updates are consistent with code changes.
Spawn two review subagents in parallel, each with a distinct perspective. A common failure mode is reviewing changes in isolation — one agent focuses narrowly on the changed files while missing how those changes interact with the rest of the project. These two agents address that by splitting the review into complementary scopes.
Focus: the changed code itself.
CLAUDE.md by walking up from each changed file's directory to the repo root.CLAUDE.md/AGENTS.md guidance.Focus: how the changes fit into the broader project.
After both subagents complete:
If any issues are found:
Skip deep quality review for binary/generated files.
Commit only after Phase 2 documentation updates and Phase 3 review are both complete and clean. Group changes into topical commits. Never combine unrelated changes.
<type>: <concise description>
<optional body explaining why, not what>
Co-Authored-By: Claude <model> <noreply@anthropic.com>
Types: feat, fix, refactor, docs, test, chore
For each topic group:
git add <specific files>)git status after each commitdocs: update project documentation