Agent Skill Pack: 24 SKILL.md Workflows for Claude Code, Cursor & Codex (Plan, Review, Test, Debug, Release + Installer & Linter)

Stop pasting the same "review this properly" instructions into your agent every day.

This pack gives your coding agent 24 task workflows ("skills") it loads only when the job comes up: a plan before code, a ranked PR review, a flaky-test hunt that fixes the cause instead of adding retries, a release checklist that stops before publishing, a dev log written from your real commits. Each skill is a folder with a SKILL.md in the open Agent Skills format, so the same files work in Claude Code, Cursor (2.4+) and OpenAI Codex. Call one by name (/pr-review) or just describe the task and let the agent pick it.

Rules vs skills. Rules (AGENTS.md, CLAUDE.md, Cursor rules) are always loaded: project facts and conventions. Skills are on-demand procedures: steps, checklists and output formats for one kind of task. Agents see only each skill's name and description until they use it, so installing all 24 costs little context. This pack pairs with Agent Rules Kit from the same store (the always-on half); each works on its own.

The 24 skills

- Planning & multi-agent: spec-to-plan, task-breakdown, worktree-split (one git worktree and branch per agent, non-overlapping file scopes, merge order), agent-handoff (verified status, decisions, dead ends, next actions), repo-onboarding

- Review: pr-review (must fix / should fix / nit, with file:line and the checks actually run), security-audit (trace untrusted input, auth, secrets, injection, SSRF, dependencies; exploit scenario per finding)

- Refactor: safe-refactor (tests green before and after every step), dead-code-cleanup (delete only with evidence), dependency-upgrade (read changelogs, one major at a time)

- Testing: write-tests (in your project's style, and proven able to fail), fix-flaky-test, tdd-loop

- Debugging: bug-repro-root-cause (repro first, git bisect, regression test), perf-investigation (measure, profile, fix the top hotspot, re-measure)

- Release: changelog-update, release-checklist (go/no-go; prepares tag/publish commands but waits for your yes), version-bump

- Docs: readme-writer (every command verified or flagged), adr-writer

- Git: commit-message (follows your repo's convention), pr-description (uses your PR template)

- Build in public: devlog-from-commits (a day's commits across repos into an honest dev log), build-in-public-thread (X/Bluesky/LinkedIn drafts with character counts; drafts only, never posts)

Every skill has a "when to use" description, numbered steps with verification, a checklist or rules, an output format, stop-and-ask conditions and a definition of done. They tell the agent not to invent numbers or claim tests it didn't run.

Also included

- install.py: install all skills, a category, or a few by name into a project or your user folder, for Claude Code (.claude/skills), Codex and Cursor (.agents/skills) or Cursor (.cursor/skills). --list, --dry-run, and it never overwrites existing skills without --force (old copies go to a backup folder outside the skills folder).

- lint_skills.py: checks any skills folder, including skills you write yourself: YAML that won't parse, invalid or reserved names, names that don't match the folder, missing or vague descriptions, over-long fields, leftover placeholders, missing referenced files. --json and --strict for CI.

- Prompt versions of all 24, generated from the same files: Cursor custom commands (.cursor/commands) and plain paste-anywhere prompts.

- 4 helper scripts (print or read only): worktree command planner, flaky-test repeat runner, semver calculator, commit collector. 6 templates/checklists: handoff note, review severity guide, security checklist, changelog, README skeleton, ADR.

- Docs: where each tool loads skills (with links to each tool's docs), a guide to writing your own skills, and a 3-page Quick-Start PDF. Unit tests for the tools.

How it works

1. python3 tools/install.py --list

2. python3 tools/install.py --all --project your-repo (add --tool agents for Codex/Cursor)

3. In your agent: /pr-review, or "review my branch".

Compatibility. SKILL.md frontmatter uses only fields from the open Agent Skills specification (name, description, metadata), following the formats documented by Claude Code, Cursor and OpenAI Codex as of November 2026; those are also the only fields claude.ai skill uploads accept. Tools without skills support can use the prompt versions. Scripts need Python 3.8+ (standard library only) and git, on macOS, Linux or Windows. Everything is plain Markdown you can edit.

Honest limits. Skills guide agents; they don't enforce anything and no workflow guarantees a particular result. Review what your agent does, especially commits, releases and anything public. Whether an agent auto-selects a skill depends on the tool and model; calling it by name is the reliable way. These are general workflows: add your own commands and conventions for best results. Tool behavior changes often, so the docs link each source for you to re-check.

License. Use in any number of your own, your employer's or your clients' projects, commercial or not, and commit the files to those repos. Don't resell or redistribute the pack itself.

Digital download (ZIP). Not affiliated with or endorsed by Anthropic, Anysphere (Cursor), OpenAI, or any tool named. Product names are used only to describe compatibility.