myAgent

Four role-based review skills for Claude Code (and any agent that reads SKILL.md), one always-on behavioral baseline, plus a minimal, fully documented permission harness.

Two layers, deliberately kept separate:

Always loaded — .claude/CLAUDE.md: baseline behavior for every task. Think before coding, minimum necessary changes, surgical edits, verifiable goals. This is the floor everything else stands on.

Conditionally loaded, four skills, that fire only when relevant:

skill	fires when
senior-fullstack-review	building/reviewing a feature that spans UI and API
senior-backend-design	designing/reviewing endpoints, schemas, jobs, service boundaries
ai-engineering-review	building/reviewing anything with an LLM in the request path
ai-eval-design	testing or measuring the quality of an AI feature

The skills assume CLAUDE.md's behaviors are already happening and add domain-specific things to check for on top.

On-demand — .claude/agents/myAgent.md: an adversarial subagent. It reads the diff, checks it against 11 known ways agents fake "done" (weakened tests, swallowed errors, placeholder code, scope creep, untested claims, etc.), and returns strict pass/fail JSON, no fixes proposed, no encouragement, no partial credit for one real violation. Runs on Haiku by default to keep it cheap enough to call often.

Wired to fire automatically — the Stop hook in settings.json: every time Claude Code is about to end a turn, an agent-type hook spawns a fresh subagent that reads myAgent.md and applies its checklist to whatever's currently uncommitted. If it fails, the turn is blocked and Claude has to address the violations before it can actually stop; if it passes (or there's nothing uncommitted to check), the turn ends normally.

Two things to know about this before you rely on it:

• It's genuinely automatic now, not opt-in. Every Stop event costs one subagent call. For a long back-and-forth session that's a lot of small checks; if that's too chatty or too expensive for your workflow, delete the Stop block from settings.json and go back to invoking myAgent by name only.
• Agent-type hooks are marked experimental in Anthropic's own documentation — the mechanism may change in a future Claude Code release. There's a built-in safety net regardless (Claude Code caps consecutive Stop-hook blocks at 8 by default and ends the session with a warning rather than looping forever), but test this in your own environment before trusting it on something high-stakes. Run a throwaway change through it and confirm it actually blocks on an obvious violation (e.g. delete a test's assertion and see if it catches it) before assuming it's load-bearing.

What actually happens when you install this

Read install.sh It downloads this repo's tarball from GitHub, and copies these into your project:

• skills/* → .claude/skills/
• .claude/settings.json → .claude/settings.json
• .claude/CLAUDE.md → .claude/CLAUDE.md
• .claude/agents/myAgent.md → .claude/agents/myAgent.md
• MEMORY.md → MEMORY.md

Nothing else. No .mcp.json, no loop-runner, no third-party servers; those are opt-in additions you bolt on yourself if you want them (see below), not defaults.

No files outside the destination directory are touched. No network calls happen other than the one GitHub download.

Recommended: clone and read, don't pipe-to-bash

git clone <this-repo-url>
cat install.sh                      # it's short, actually read it
cd your-project
DEST=. bash /path/to/myAgent/install.sh

If you want the one-liner

curl -fsSL <raw-url>/install.sh | bash

Existing files are skipped by default (never silently overwritten). Pass FORCE=1 to overwrite. Pin to a specific release with REF=<tag> instead of trusting whatever is currently on main.

The permission model — read this before your first session

.claude/settings.json is the actual trust boundary: it decides what the agent can do without asking you. This kit ships with:

Allowed without prompting: read-only and low-risk git operations (status, diff, log, add, commit, branch, checkout), and common test/lint/build commands (npm test, npm run lint, npm run build, pytest, ruff, mypy).

Explicitly denied, even if the agent tries: force-pushes, rm -rf, curl/wget (no arbitrary network fetches from the shell), sudo, chmod 777, and anything piping into bash/sh.

Anything not on the allow list falls back to Claude Code's normal ask-first behavior. Adjust this list for your own workflow — it's a starting point, not a claim that it's right for every project. If you add a rule, add a one-line comment in your fork's README explaining why, the same way this one does.

Optional add-ons (not installed by default)

If you want the MCP wiring or the run.sh loop-runner pattern from the original loopkit project, add them yourself deliberately:

• MCP servers: only wire up servers you've personally checked. Pull tokens from environment variables, never commit them.
• Loop runner: a while true script that calls claude -p repeatedly is easy to write in 8 lines if you want unattended runs — but it means the agent is acting without a human in the loop each turn, so it deserves a tighter settings.json than the interactive default above, not the same one.

Writing your own skill

Each skill is one file: skills/<name>/SKILL.md with YAML frontmatter (name, description) and a body.

The description is what triggers auto-loading — be specific about when it should and shouldn't fire, not just a topic label. Vague descriptions cause skills to either never trigger or trigger on the wrong tasks.

Keep the body a checklist of what to look for and common failure patterns, not a tutorial. The model already knows how to code; the skill's job is to point it at the specific things it tends to skip.

Shoutouts

https://github.com/Archive228/loopkit

License

MIT. Customize as needed.