claude-code-toolkit

Complete the inner development loop in one command: stage changes, write a conventional commit message, push, and open a pull request with a structured description. Uses the commit-message and pr-description skills. Inspired by Boris Cherny's workflow — he runs this dozens of times daily.

Critically review a plan before execution. Spawns a subagent that acts as a skeptical staff/distinguished engineer poking holes in the current plan — finding missed edge cases, incorrect assumptions, scope risks, and architectural concerns. Use after Plan Mode produces a plan and before switching to auto-accept.

Encode a mistake into a permanent improvement. After Claude makes an error, run /sharpen to create a test, lint rule, or CLAUDE.md entry that prevents it from recurring. This is the sword-sharpening loop — each session makes the next one more capable.

Audit .claude directory files (CLAUDE.md, rules, docs) for factual drift from the current codebase. Produces a report and applies corrections only with explicit human approval. Never modifies commands, skills, agents, architectural decisions, or workflow rules. Use after significant refactors, dependency changes, or new features. Run weekly.

Comprehensive code review using parallel subagents. Reviews recently changed files for bugs, security issues, performance problems, and convention violations. Use before committing, before PRs, or after implementing a feature. Optionally, pass a focus area like "review auth changes" or "review the new API endpoint implementation".

Audit two related repositories (e.g., frontend and backend, two microservices, a library and its consumer) for consistency in documentation, .claude configs, API contracts, shared types, and architectural assumptions. Use when you suspect drift between repos or after major changes to either side. Read-only — produces a report, never modifies.

Investigate bugs, errors, and unexpected behavior. Traces data flow, analyzes error messages, forms hypotheses, and proposes fixes. Use when something is broken and you need to understand why. Provide an error message, a description of unexpected behavior, or a failing test.

Audit and update project documentation to match current code. Use when docs may have drifted from implementation — after refactors, new features, or dependency upgrades. Covers READMEs, onboarding guides, architecture docs, API docs, inline JSDoc/GoDoc, and any markdown in the repo. Does NOT cover .claude/ directory. That is claude-sync's job.

Generate git conventional commit messages from staged changes or diffs. Produces structured commit messages with type, optional scope, subject, and body. Use when committing code or when asked to write a commit message.

Maintain task-specific context files (plan, context, tasks) that persist across Claude Code sessions. Use when starting a non-trivial feature, when resuming work from a previous session, or when asked to create or update development task documentation. Bridges the "daily amnesia" problem.

Generate structured pull request descriptions from diffs or branch changes. Produces a clear summary, change list, testing notes, and review guidance. Use when creating a PR or when asked to write a PR description.