mistakenot
UserComposable tools for autonomous coding with Codex and Claude.
Categories
Indexed Skills (32)
contextual-commit
Write contextual commits that capture intent, decisions, and constraints alongside code changes. Use when committing code, finishing a task, or when the user asks to commit. Extends Conventional Commits with structured action lines in the commit body that preserve WHY code was written, not just WHAT changed.
new-package
Scaffold a new auto-* package in the auto-stack monorepo. Creates the full directory structure, Go source files, go.mod, CLAUDE.md, and Makefile integration following established patterns. Use when: "create a new package", "add auto-foo", "scaffold a new tool", "new auto package".
recall
Reconstruct and narrate the current development context from contextual commits. Run at session start, when resuming work, or when switching branches. Produces a brief, conversational summary of where things stand.
release
Create a new auto-stack release by tagging a commit and pushing the tag. The GitHub Actions release workflow builds binaries and publishes the release.
feature-miner
This skill should be used when the user asks to "mine sessions for feature ideas", "find usage patterns in coding sessions", "discover autodoc improvements from real usage", "run feature miner", "mine cass for ideas", "what are agents struggling with in docs", or mentions mining coding history for autodoc improvements.
readme-updater
Bring a project's README up to date with the current state of the codebase. Audits the existing README for stale commands, dead links, renamed tools, and drifted status claims; surveys what's changed since the README was last touched; rewrites it with verified facts; then commits and pushes. Use this skill whenever the user says "update the README", "the readme is stale", "refresh the README", "rewrite README.md", "add new features to the readme", or otherwise asks for documentation that reflects the current state of the project. Also trigger when a README clearly hasn't kept pace with the codebase (e.g. recent feature commits not reflected, broken examples).
self-improve
Run the self-improvement loop on a focus area. Use when the user asks to improve, fix recurring problems, or run self-improve on a tool. Trigger when: "self-improve", "improve autosearch", "find and fix problems in". Do not trigger for doc reviews or one-off bug fixes.
address-feedback
Works through all open PR review threads by fixing code, replying to reviewers, and resolving threads. Use when 'address feedback', 'fix PR comments', 'resolve review threads', or after a code review has left feedback on the PR.
cass
Coding Agent Session Search - unified CLI/TUI to index and search local coding agent history from Claude Code, Codex, Gemini, Cursor, Aider, ChatGPT, Pi-Agent, Factory, and more. Purpose-built for AI agent consumption with robot mode.
code-review
Performs a structured code review of changes with severity labels covering correctness, security, performance, error handling, testing, and architecture. Use when 'code review', 'review this PR', 'review changes', or when a structured review of implementation changes is needed.
commit-task
Verifies completeness of all planning docs and commits them to main. Use when 'commit the task', 'finalize the task', 'commit task docs', or after all planning docs have been reviewed and approved. Not applicable when implementation has already started or for committing code changes.
execute-task
Autonomously implements a planned task end-to-end using the coordinator-subagent pattern. Reads task docs, creates a worktree, dispatches subagents per phase, tracks progress, and opens a PR. Use when 'execute task', 'implement task', or after planning docs have been committed.
generate-10-ideas
Generate 10 brilliant ideas for powerful new functionality. Use when brainstorming features, improvements, or innovations for a system. Internally generates 100 candidates and filters to the top 10 most impactful, pragmatic, and innovative ideas.
improve-claude-md
improve a CLAUDE.md file using <important if> blocks to improve instruction adherence
new-plan
Writes context.md and plan.md for an existing task by gathering codebase context and breaking the solution into executable phases. Use when 'write a plan', 'new plan', 'create the plan', 'plan the task', or after solution.md has been approved. Not applicable when solution.md doesn't exist yet.
new-solution
Writes a solution design (solution.md) for an existing task by exploring approaches and tradeoffs. Use when 'write a solution', 'design the solution', 'new solution', 'explore approaches', or after requirements have been approved. Not applicable when requirements.md doesn't exist yet.
new-task
Creates a new task folder with requirements.md from user input. Use when 'create a task', 'new task', 'start a task', 'write requirements', or when a user describes a feature/fix to plan. Not applicable when executing an existing task or reviewing existing docs.
open-prose
OpenProse is a programming language for AI sessions. Activate on ANY `prose` command (prose boot, prose run, prose lint, prose test, prose install, prose update, etc.), running .prose or .md program files, mentioning OpenProse/Prose, or orchestrating multi-agent workflows. The skill intelligently interprets what the user wants.
reflect-on-agent-sessions
Reflect over coding agent sessions using autosearch, find recurring workflow problems, and produce a markdown report with reproducible query evidence and explicit reasoning for each issue.
request-codex-review
Runs a code review of task planning docs via Codex, then resolves any comments left. Use when 'request codex review', 'get codex review', 'codex review task', 'have codex review', or when a second pair of eyes is needed on task docs before execution.
review-resolve
Use when the user asks to "resolve comments", "address comments", "reply to comments", or "go through comments" in markdown planning docs. Resolve or respond to review threads using the project's authoritative comment format.
review-task
Reviews task planning documents (requirements.md, solution.md, context.md, plan.md) and leaves structured inline comments flagging problems and improvements. Use when 'review task docs', 'review the plan', 'review task 042', 'check the planning docs', or a task ID/folder for doc review. Not applicable for code review of implementation changes.
skill-reviewer
Review all compiled skills against multi-agent best practices and produce a prioritized report. Use when the user asks to 'review skills', 'audit skills', 'check skill quality', 'lint skills', or wants a quality report on the skills in this repo.
tech-spike
Run an exploratory tech spike to validate assumptions, investigate edge cases, and stress-test ideas before committing to implementation. Use when 'tech spike', 'spike this', 'explore whether', 'validate assumptions', 'investigate feasibility', 'proof of concept', 'is this possible', 'what are the risks', 'test this approach', or when the user describes an idea they want to de-risk before building. Also use when requirements seem uncertain and the user wants evidence before committing to a design.
experiment-designer
Design and run a formal hypothesis-driven experiment end-to-end — write the design doc with go/no-go criteria, dispatch workers, validate their output, and capture findings. Use when 'design an experiment', 'experiment-designer', 'run an experiment to test', 'set up a hypothesis test', 'fan out experiments to figure out', or when the user wants a structured multi-phase investigation rather than a quick exploration. Distinct from tech-spike, which is for fast informal exploration; use experiment-designer when the question deserves a written hypothesis, explicit pass/fail thresholds per metric, and findings that get committed to git as long-lived artifacts.
request-claude-review
Runs a code review of task planning docs via Claude Code, then resolves any comments left. Use when 'request claude review', 'get claude review', 'claude review task', 'have claude review', or when a second pair of eyes is needed on task docs before execution.
complete-task
Finalizes a feature branch and merges it to main. Addresses remaining PR feedback, runs tests, writes feedback.md, merges the PR, exits the worktree, and verifies the merge. Use when 'complete task', 'finalize and merge', 'ship this task', or when the feature branch is ready to ship.
delegate-task
Dispatches task execution to an idle Claude Code pane in a tmux session, freeing the current session for other work. Use when 'delegate task', 'send to executor', or when the user wants to hand off a task to a background pane.
executor-status-check
Monitors all executor panes in a tmux session and reports the status of running, completed, stuck, or idle tasks. Use when 'check executors', 'executor status', 'how are tasks going', or when monitoring background task progress.
resolve-comments
Resolves inline review comment threads in markdown planning docs by fixing content, rejecting invalid concerns, or escalating blockers. Use when 'resolve comments', 'address comments', 'fix review comments', 'go through comments', or after a review has left comments on task docs. Not applicable for PR code review feedback (use address-feedback instead).
revise-readme
Update README and documentation to reflect all recent changes. Use when new features, commands, or options have been added and docs need to catch up. Frames everything as current state — never references what changed.
task-feedback-analyser
Extracts recurring patterns from completed task feedback into generalizable workflow rules. Scans feedback.md files and review threads, clusters by theme, and drafts rules with a strict 3-example minimum. Use when 'analyse feedback', 'extract rules', 'find patterns in feedback', or when reviewing historical task outcomes.
Bio shown is the top-scored skill's repo description as a fallback — real GitHub bios land in a future update.