Reusable Agent Skills for PRs, review, and shipping. Works with Cursor and other Agent Skills hosts.
Claude Code reads AGENTS.md for install and conventions when CLAUDE.md is absent. That does not load the always-on rules. Those stay in rules/*.mdc. Copy or point at them (see How vendors consume the kit).
Install all skills:
npx skills add netanelavr/agent-kit --allInstall a single skill:
npx skills add netanelavr/agent-kit/skills/kit-create-prAlways-on rules are not installed by npx skills add. Copy rules/*.mdc if you want the always-on kit loop. Skills install does not do that. See How vendors consume the kit.
| Skill | When to Use |
|---|---|
kit-create-pr |
Commit changes and create a pull request with proper formatting |
kit-triage-comments |
Triage PR comments — batch or one-by-one before fixing |
kit-review |
Author self-check or reviewer second-pass (author / reviewer) |
kit-optimize-context |
Reduce always-on context across Cursor / Claude / Codex / etc. |
kit-debug-prod |
Production bugs via logs/traces/errors/metrics (needs team wire-observability.md) |
kit-challenge-design |
Stress-test a plan or design against domain language, docs, and code |
kit-plan |
Align on a buildable concept (includes three readings when the ask is fuzzy); wait for approval |
kit-decisions |
Living log for one large feature: accepted, rejected, landmines, open |
kit-minimize |
Smallest useful path / minimal PR (not design grill) |
kit-strip-diff |
After the outcome works: drop unused additions, extra instruction lines, and format-only hunks |
kit-prove |
Live proof for “it works” claims — smoke or runtime path; probes stay temporary |
kit-explain |
Problem / Solution / How + one compact mermaid |
kit-learn-repo |
Ramp on a repo: layout, run/test, safe change points |
kit-check-blast |
What else could this change break — prove the safety fact |
kit-audit-metrics |
Honest aggregate analytics — define, query, provenance, no invented numbers |
kit-audit-skills |
After adding/renaming skills: overlaps, contradictions, stale names, catalog drift |
kit-capture-lesson |
Turn a steering/repeated miss into a short written rule; wait for approval |
Always-on rules (rules/*.mdc). Copy or point at them; do not rely on slash invoke. See How vendors consume the kit.
| Rule | When it applies |
|---|---|
kit-comment-why |
WHY/landmine comments only — no WHAT narration |
kit-delegate |
Inline vs skill vs subagent; safe parallel fan-out |
kit-keep-diff |
always — minimal diffs, no unrelated reorders |
kit-humanize |
always — plain writing; no AI-slop in commits/PRs/docs |
kit-work-loop |
always — plan first, isolate noisy work, strip a fat diff, isolated prove, capture repeated fixes |
Invoke a skill in Cursor with /<skill-name>.
Skills install the same way on every host: npx skills add. Rules are separate.
| Vendor | Skills | Rules |
|---|---|---|
| All Agent Skills hosts | npx skills add netanelavr/agent-kit --all (or a single skill path) |
not installed by the CLI |
| Cursor | same | copy rules/*.mdc into .cursor/rules/ (or user rules) |
| Claude Code | same | copy rules/*.mdc into the consuming repo, or point at them from root AGENTS.md. Claude Code reads AGENTS.md as a fallback only when CLAUDE.md is absent. If you already have a CLAUDE.md, delete it or reference AGENTS.md / the .mdc files from it |
| Codex | same | copy rules/*.mdc, or point at them from root AGENTS.md |
MIT