Three specialist subagents and three skills for Claude Code: code review, debugging, planning, and the discipline around them.
These are prompt files, not magic. They plug into Claude Code's native
subagent (.claude/agents/) and skill (.claude/skills/) machinery and make
it do disciplined work: reviews that hunt outside the diff, debugging that
proves the cause before fixing, "done" that requires evidence. Each file is
written to the same bar: concrete method, failure-mode table, definition of
done. The recorded sessions below are what that produced on real code.
git clone https://github.com/Honorboxx/crew && cd crew
sh install.shThe installer symlinks the pack into ~/.claude (so git pull
updates it), tracks everything it installs in a manifest, and will not touch
files it doesn't own. sh install.sh --dry-run shows the plan first;
--copy detaches from the checkout; --uninstall removes exactly what it
installed; --target some/project/.claude installs per-project instead.
POSIX sh, zero dependencies.
Everything is crew-prefixed so it never collides with your other packs.
crew-reviewer: correctness pass. Reads your diff three times: intent,
adversarial, and outside the diff. Greps callers of everything you changed,
because most review-detectable bugs are in unchanged code whose assumptions
just broke. Findings ranked P0/P1/P2 with a concrete failing scenario each,
plus an explicit "checked and clean" list. Won't manufacture nits.
Recorded session (real, verbatim-condensed): crew-reviewer on a shipped commit of the store this pack funds, a markdown image/gallery feature the author had marked "(tested)". It found two P1s and a P2 in production code; all three were fixed the same hour.
crew-reviewer on 936ae3b "theme preview gallery" — verdict: fix before relying on it. Ran the suite (13/13 green) and the build (clean), then drove renderMarkdown with adversarial input. P1 scripts/lib/md.js:51 — image src interpolated raw into <img src="..."> while every other URL sink in the file is escaped. Demonstrated: ![x](/a.png"onerror="alert(1) → <img src="/a.png"onerror="alert(1"...> — an injected handler. The scheme allowlist doesn't help: the payload starts with "/". Fix: escapeHtml the URL like link hrefs already do. P1 scripts/test/core.test.js:144 — the safety test is decoration: its payload has a paren in the URL, so it fails the image-line regex and never reaches the new filter; delete the filter and the test still passes. Fix: paren-free unsafe src that actually hits the code path. P2 md.js:47-54 — an image run whose srcs are all rejected vanishes silently: green build, missing content, zero signal. Checked and clean: both renderMarkdown callers wrap output in .prose so the new CSS applies (verified in dist/); 4 consecutive images group into one gallery, single → figure; block ordering doesn't mis-swallow paragraphs; alt text IS escaped; assets/ passthrough resolves; build idempotent; 13/13 existing tests meaningful except the one above.
crew-debugger: root cause with receipts. Reproduce → shrink → ranked
hypotheses → experiments that eliminate rather than confirm → fix → prove
(repro fails before, passes after, both outputs shown) → sweep for sibling
bugs. Keeps a hypothesis log so long sessions never re-test eliminated ideas.
Refuses to ship a fix without a causal chain.
Recorded session (real, verbatim-condensed): crew-debugger on a real production incident, a fulfillment failure that got an order permanently stuck. No live APIs: it ran the real engine as a child process with only the network faked, replaying the verbatim production error.
Incident: fulfillment logged FAILED ... invite <owner> -> 422 "Repository owner cannot be a collaborator" when the store owner test-bought their own product. The catch path marked the session processed → TERMINAL: ledger needs_attention, no retry. Repro (decision path, not HTTP): real fulfill.js as a child process, global.fetch stubbed to documented PUT-collaborator semantics + the verbatim 422 body. Symptom reproduced byte-identically. Hypotheses: H1 wrong username/repo plumbing — ELIMINATED (prod state shows correct field + grant flowed through; engine did the right thing) H2 no owner-self-purchase guard — CONFIRMED (invites unconditionally) H3 token/API misuse — ELIMINATED (422-custom, not 401/403; docs confirm request shape + 201/204 handling correct) Fix: isRepoOwner() guard — owner buys → fulfilled, no invite. Regression test fails before, passes after; control test pins normal buyers to 1 invite. Sibling sweep: refund bot already guards the same case; a dead duplicate core file (drift hazard) flagged for deletion; the deeper debt — ANY failure is terminal because processed.push runs in the catch path, so a transient 5xx never retries. Named for a deliberate decision.
crew-planner: options, then a plan. For tasks where the approach itself
is a decision. Recon in the codebase first, 2-3 genuinely different options
with the rejected one and why, then steps sized by risk (every step with an
observable done-check), and the section that matters most: explicit non-goals.
crew-verify: evidence before claims. The gate between finishing work
and describing work. Turns each claim ("fixed", "passing", "works") into the
cheapest observation that would prove it, made this session, and includes
the trick most setups miss: verify the verifier (green that can't go red is
not evidence). Then it runs the other half nobody runs: the claims your change
just falsified somewhere else, the counts and READMEs and screenshots no test
covers. Anything unobserved ships labeled "changed, not verified".
crew-scope: size first, then plan that much. S/M/L sizing by counting
plausible designs, a 5-line mini-plan for M tasks (every line objection-bait),
dispatch the planner for L. Kills both failure modes: coding an L blind, and
planning theater on an S. Plus the mid-task rule for "while I'm here" creep:
park it, don't chase it.
crew-git: history as a communication act. Atomic commits tested by
"can this revert alone", the staged-diff read-through (the medium switch that
catches debug prints), why-not-what messages, and a safe-repair table,
including the rule that matters when it matters: a leaked secret is rotated
first; history rewriting is only tidying.
Every transcript in this README is a recorded session. There are no illustrative examples here, because an invented transcript proves only that we can write one. The three skills above ship without a sample for an honest reason: a skill is a discipline you invoke inline rather than an agent that reports, so there is no artifact to record that would not be staged.
Two further recorded runs, including a paid agent and the free reviewer working the same public commit, are in SAMPLES.md.
This repo is the free tier, and it's genuinely useful alone. The full pack is the whole system: ten agents and fourteen skills that cross-reference each other, plus hooks and starter templates.
Before deciding, read SAMPLES.md. A paid prompt pack is the
one kind of product you cannot evaluate before paying, so that page opens it:
the complete text of one paid agent and one paid skill, the description of
all eighteen paid-only files, the exact diff between the six free files and
their paid counterparts, a recorded run of a free and a paid agent on the same
public commit (including the live bug it found), and our own account of the
three weakest files in the pack.
| Crew (this repo) | Crew Full ($19) | |
|---|---|---|
| Agents | 3: reviewer, debugger, planner | 10: + simplifier (second review pass), tester, perf, security, refactorer, docs, release captain |
| Skills | 3: verify, scope, git | 14: + TDD, debug loop, self-review, PR authoring, changelog, release, docs pass, refactor, perf triage, security pass, handoff |
| Hook examples | none | 6, safe and commented (git, shell, exfil, scope and secret guards, plus format-on-edit), with their test suites |
| CLAUDE.md starter templates | none | solo + team variants |
| Design doc | none | ROSTER.md: every agent/skill mapped to a job and a boundary |
| Installer | this one | same one, covering the full pack |
| Updates | git pull |
private repo access; updates land there |
| License | MIT | per-developer commercial |
Buying replaces the six free files with superset versions (added cross-references into the full system). The installer handles it cleanly, and the exact diff is published so you can see precisely what changes before you pay.
- These improve discipline, not model capability. A model that can't find a bug won't find it because a prompt told it to be systematic, but it will stop claiming "fixed" without running the repro.
- Agents inherit your configured model and burn tokens like any subagent; the reviewer pass on a big diff is not free.
- Written for Claude Code's agent/skill format. Other tools may read the markdown fine, but the dispatch/invoke mechanics are Claude Code's.
MIT. See LICENSE.