Claude Code ships with no opinion about your repo. This is the opinion: 10 skills for the work you repeat, 8 hooks for the mistakes that cost the most. Here is the whole file list and two of the files, so you can decide you do not need it.
Last updated: 2026-09-22
Every line below was captured from the pack’s own hooks and test suite — the block messages are the hooks’ real wording, and the tally is what that suite really printed on the day it was recorded (it is 95 now) — real last line. Nothing is a mockup.
Three of the eight hooks are shown. The other five: protect-branch.sh, no-secret-reads.sh, format-changed.sh, big-diff-warn.sh, session-summary.sh. Two of the files are printed in full further down.
A pack of “skills and hooks” is not something you can judge. Nothing below is a surprise on opening.
| Path | What it does |
|---|---|
| 8 hooks — they run on Claude Code events, not on a schedule | |
hooks/block-secrets.sh | Refuses a commit carrying a credential |
hooks/no-secret-reads.sh | Refuses reading a .env or key file into the transcript |
hooks/no-destructive-shell.sh | Stops rm -rf and git clean in their many spellings |
hooks/protect-branch.sh | Stops a commit landing straight on main |
hooks/no-skip-tests.sh | Refuses the flags that make a test run lie — printed in full below |
hooks/big-diff-warn.sh | Warns when a branch has grown past reviewable |
hooks/format-changed.sh | Formats what changed, not the repo |
hooks/session-summary.sh | Writes what the session did on exit |
| 10 skills — you run them by name | |
skills/pr-description/ | PR text from the real diff — printed in full below |
skills/test-first/ | Failing test before the fix |
skills/debug-loop/ | Reproduce, narrow, fix, prove |
skills/security-sweep/ | A pass over what you are about to ship |
skills/review-mine/ | Review your own branch before a human does |
skills/refactor-safe/ | Change shape without changing behaviour |
skills/flaky-hunt/ | Find the test that only sometimes fails |
skills/dep-check/ | What a dependency bump actually touches |
skills/onboard-repo/ | Get oriented in a repo you did not write |
skills/release-notes/ | Notes from the commits, not from memory |
| 10 more | |
.claude/settings.json | Wires all 8 hooks to their events |
test/run-hooks-test.sh | 134 offline checks, ~2s. Run it before you trust any of this |
test/run-skills-test.sh | 42 offline checks, ~1s. Proves all ten skills actually load — a skill with broken frontmatter fails silently |
test/mutate-skills-suite.sh | Deletes each check in the skills suite in turn and proves the suite goes red. 16 checks, 16 caught. A passing suite proves the checks ran, not that any of them matter |
test/mutate-hooks-suite.sh | The same proof for the hooks suite: breaks each blocked pattern in turn and proves a check goes red for it. 81 patterns, 81 caught, 0 survived |
CHECKLIST.md | Ten minutes, once, proving each hook really fires |
INSTALL.md | Merging into a .claude/ you already have — the common case |
README.md, CHANGELOG.md, LICENSE | The usual |
No dependencies, no network, no telemetry. macOS and Linux.
.claude/hooks/no-skip-tests.shThe whole file, exactly as it ships. It blocks a Bash call that runs tests with a flag that makes the run report success no matter what happened.
#!/usr/bin/env bash
# PreToolUse(Bash): refuse the flags that make a test run lie.
set -uo pipefail
payload="$(cat)"
JQ="${JQ_BIN:-jq}" # override to test the no-jq fallback: JQ_BIN=/nonexistent
if command -v "$JQ" >/dev/null 2>&1; then
cmd="$(printf '%s' "$payload" | "$JQ" -r '.tool_input.command // empty' 2>/dev/null)"
else
cmd="$payload"
fi
[ -z "$cmd" ] && exit 0
case "$cmd" in
*test*|*pytest*|*vitest*|*jest*|*go\ test*) ;;
*) exit 0 ;;
esac
bad="$(printf '%s' "$cmd" | grep -oEm1 -- '(--passWithNoTests|--no-verify|--exitcode=0|--ignore-errors|\|\| *true|; *true$)' || true)"
if [ -n "$bad" ]; then
echo "BLOCKED by .claude/hooks/no-skip-tests.sh: '$bad' makes this run report success whatever happens." >&2
echo "Run the tests for real. If some genuinely cannot run here, say which and why instead of masking the exit code." >&2
exit 2
fi
exit 0The part worth stealing even if you never buy this: exit 2, not exit 1. Claude Code treats 2 as “blocked, and here is why” and feeds stderr back; 1 is just a failure. And the message says what to do instead of masking the exit code, because a hook that only says no gets removed within a week.
.claude/skills/pr-description/SKILL.mdAlso the whole file. Skills are short on purpose; the value is in what they refuse to let you skip.
---
name: pr-description
description: Write a PR description from the actual diff — what changed, why, how it was verified, what to watch. Use before opening a PR.
---
1. `git diff main...HEAD --stat` then `git log main..HEAD --oneline`.
2. Write, in this order:
- **What** — 1–3 sentences. What behavior is different now.
- **Why** — the problem, not the solution restated.
- **How it was verified** — the commands you actually ran and their result. If you didn't run them, write "not verified" rather than implying you did.
- **Risk / what to watch** — the thing most likely to break, and where it would show.
3. Keep it under 200 words. Link issues; don't paste them.
4. Don't list every file. The diff is right there.Note the fourth bullet: “if you didn’t run them, write ‘not verified’ rather than implying you did.” That line is the whole reason the skill exists.
They are more of exactly this: short, offline, readable in a sitting. There is no framework in the zip. If what you just read looks like something you would write yourself in an afternoon, you are right, and you should — the shape is not a secret and this page is most of it.
What the pack is not: it is not a setup service, it does not know your repo, and nobody runs it for you. Files land in .claude/ and do nothing until you use them. If you want this written from your code, with every command run in your repo first, that is Setup Lite $29 — a different purchase and it exists today.
28 files, one zip, download link on the receipt page. You just read two of them in full and the file list is above — that is the product, not a teaser of it. No email wall, no account, no subscription.
Buy the Skills & Hooks Pack $29
Not sure it is worth paying for at all? The free starter repo is public, complete, and the same shape — generic on purpose. Take that first. We would rather you had it than nothing.
28 files in one zip: 8 hooks wired to Claude Code events, 10 skills you run by name, a .claude/settings.json that wires the hooks to their events, two offline test suites totalling 176 checks, plus a mutation harness that proves those checks are load-bearing, CHECKLIST.md, INSTALL.md, and README, CHANGELOG and LICENSE. $29, downloaded from the receipt page. The complete file list is on this page and two of the files are published here in full.
No. No dependencies, no network and no telemetry, on macOS and Linux. The two suites, test/run-hooks-test.sh and test/run-skills-test.sh, run offline in about three seconds and should end 134 passed, 0 failed and 42 passed, 0 failed — run it before you install anything.
Claude Code treats exit 2 as “blocked, and here is why” and feeds stderr back to the model; exit 1 is just a failure. A blocking hook should also say what to do instead of masking the problem, because a hook that only says no gets removed within a week.
.claude/ directory I already have?Yes, and that is the common case. INSTALL.md covers merging rather than copying over what is already there. It takes about two minutes.
The pack is generic files you install yourself: it does not know your repo and nobody runs it for you. Setup Lite $29 is the same shape written from your code — a CLAUDE.md, a permission allowlist and one skill — handed back as a pull request in about 24 hours.
Yes. The kit-claude-code-starter repo is public, complete and the same shape, generic on purpose. Take that first if you are not sure you need the pack.