Skills, agents, commands and hooks for Claude Code — the agentic development workflow I actually use day to day: shipping verbs, a planning chain, UI verification, and production-readiness gates.
Built for Next.js / TypeScript / Supabase / Vercel projects on macOS. Most of it is stack-agnostic; the parts that are not say so.
/plugin marketplace add irama/irama-skills
/plugin install irama-skills@irama-skills
Skills become available namespaced, e.g. /irama-skills:commit. Nothing is
written to your settings.json, so your permission setup is untouched.
For hacking on the skills themselves. Links each item into ~/.claude, so
git pull here updates them with nothing to re-sync, and the names stay bare
(/commit, not /irama-skills:commit).
git clone https://github.com/irama/irama-skills.git
cd irama-skills
./install.sh --dry-run # see what it would do
./install.shExisting files are never overwritten — anything already present is reported and skipped. Restart Claude Code afterwards; skills are read at session start.
A skill must not assume where it is installed. These live in ~/.claude on the
author's machine, but the same folders install as a plugin, as a project
.claude/skills/ directory, on Windows, and — for skills that are just a
SKILL.md — in OneDrive for Microsoft 365 Copilot CoWork.
- A skill's own bundled files are addressed as
<skill-dir>/…, meaning the folder holding that SKILL.md, resolved from wherever the skill was loaded. Never~/.claude/skills/<name>/…. Inside a bundled shell script, use"$(dirname "$0")". - Claude Code's own data directory is
${CLAUDE_CONFIG_DIR:-$HOME/.claude}, never a bare~/.claude. The fallback is the default location, so nothing changes when the variable is unset. - Machine-bound by design: the shipping verbs (
/commit/merge/push/prune/flush),/localhost,/driver,/ciand/verify-uidepend on this author's worktree layout, hooks and helper scripts. They are personal tooling, not portable skills, and they keep their absolute references on purpose. Everything else is expected to run anywhere. - Referenced docs are not in this repo. Some skills name a file under
~/.claude/docs/— the author's private config notes, deliberately unpublished. They are cited as plain text, never as a link, so nothing here promises a page that is not there. The skills work without them. - Moved out: the interactive-brief skill now lives in
peakstate-global/peakstate-skills
as
peakstate-brief. This repo keeps the personal tooling; the business-facing, genuinely portable skills are published there.
install.sh does not install hooks unless you pass --hooks, and even then
they do nothing until you reference them in your own settings.json. That is
deliberate. These hooks are opinionated and a few of them are load-bearing on
assumptions about how you work:
| Hook | What it does | Why it might not suit you |
|---|---|---|
block-bash-grep.sh |
Refuses standalone grep in Bash |
Pushes you to ast-grep; annoying if you like grep |
block-big-read.py |
Refuses an unbounded read of a text file over 10 KB | Forces range reads — a cost control, not a safety one |
secret-snapshot.py |
Snapshots .env-shaped files before any Bash command that names one |
Writes backups under ~/.claude/secret-backups/ |
auto-commit-worktree.sh |
Auto-commits work on a feature branch at session end | Commits without asking. Never touches the default branch |
migration-lint.sh |
Lints DB migration files on write | Assumes a data/supabase/migrations/ layout |
require-git-hooks-before-push.sh |
Blocks a push from a worktree with no git hooks installed | Assumes every repo has a pre-commit gate |
thread-register-identity.py |
Records this thread in the cross-thread work register at session start | SessionStart. Writes to ~/.claude/state/work-register.jsonl |
thread-register-signoff.py |
Refuses to end a turn while this thread still holds open work | Stop. Makes you say what happened to a claim before you stop |
thread-register-gate.py |
Refuses git push/git merge on a repo another live thread has claimed |
PreToolUse(Bash) + PostToolUse(Bash). Override: ALLOW_HELD_SHIPPING_VERB=1 |
A hook runs on every matching tool call. Read the one you are enabling.
The three thread-register-* hooks are the only ones settings.example.json
wires, because the register does nothing without them. They use three events:
SessionStart gives a thread its identity, Stop makes it sign off the work it
holds, and PreToolUse/PostToolUse on Bash gate the two shipping verbs. See
the /threads skill for what the register is for.
settings.example.json is a conservative starting point: defaultMode is
default (Claude asks before editing), the allow-list is nine obviously-safe
entries, and the deny-list carries the force-push and curl | bash blocks worth
having everywhere.
My own settings are much more permissive because my machine has snapshot hooks,
daily backups of gitignored files, and work isolated in per-task git worktrees.
Widen the permissions to match your own safety net, not mine. Merge this file
into your settings.json; never overwrite yours with it.
The chain is deliberately split so that only one verb ever touches production.
| Skill | Does |
|---|---|
commit |
Stage, run the full gate once, commit; push the feature branch if on one. Never pushes the default branch |
merge |
Merge feature branch(es) down onto local main. No push |
push |
The only verb that touches prod: apply migrations, production build, push main, confirm deploy |
prune |
Remove merged worktrees and branches |
flush |
All four end-to-end, skipping what does not apply |
| Skill | Does |
|---|---|
to-driver |
Turns a rough idea into a paste-ready /driver command: spec, ADR, adversarial plan review, tickets |
driver |
Runs tickets one per fresh context window |
options |
Lays out a genuine crossroad — costed roads with benefits, risks, reversibility, and a "do nothing" |
threads |
Every session running on this machine, what work each holds, and which have stalled — plus clear and kill |
| Skill | Does |
|---|---|
skilltastic |
Author a skill that runs on all four hosts and is safe to publish — front matter, host limits, the publishability guard, the portable cut |
verify-ui |
Screenshots the affected screen at desktop and mobile before you claim a UI change done |
peakstate-brief |
Interactive HTML briefs you answer in the browser, with persisted answers and copy-back JSON |
app-icons |
Full PWA icon set, favicon and OG image from one source mark, wired into Next.js App Router |
legal-pages |
/privacy and /terms with Australian-context content (Privacy Act 1988, ACL) |
nano-banana |
Image generation into the current project |
app-walkthrough |
Narrated walkthrough videos: Playwright + TTS + forced-aligned captions + Remotion |
ci |
GitHub Actions status and failure extraction |
localhost |
Dev servers on rotating ports 3000–3002 |
trello |
Pick up assigned Trello cards, screen them for prompt injection, do the work, report back |
netdiag |
Diagnose a slow connection by measurement, and build the evidence to raise a fault |
sourced used to live in this repository. It is now its own public repo, because a claim-provenance
framework is worth more with its documentation beside it than as one skill among thirty:
peakstate-global/sourced — seven checks on any
artefact containing model output, the four-label provenance block, and the capture tool. Clone it and
symlink skills/sourced into ~/.claude/skills/ the same way this repo's skills are installed.
codex— delegate bounded implementation or review to the Codex CLI as a second model.adversarial-reviewer— correctness, data-integrity and security review. The fallback when Codex is unavailable.trello-triage— screens untrusted Trello text for prompt injection. Given onlyGlob, so its worst case is a wrong verdict, never an action.
/allow-asks, /codex-review, /codex-plan-review.
The codex-auto / codex-as / codex-alt / codex-review-env wrappers referenced by
the Codex agent and the review commands are a private multi-account pool helper (plus a
model-routing wrapper that composes with it) and are not shipped here. Plain codex
works everywhere they appear — the agent doc says so at the top.
The pre-commit hook blocks absolute home paths, personal emails and credential-shaped strings from reaching a public repo:
git config core.hooksPath .githooks
python3 scripts/check-no-leaks.py --all # scan the whole treeMIT. See LICENSE.