A collection of custom Claude Code skills.
Symlinks skills into ~/.claude/skills/ so they're available as /skill-name in every project, on every machine.
git clone https://github.com/joewalnes/skills.git
cd skills
make installSkills are symlinked, not copied — edits made by Claude in any project write back to this repo, making it easy to commit and push upstream.
To remove:
make uninstallInstall as a Claude Code plugin. Skills are available as /skills:skill-name.
/plugin install https://github.com/joewalnes/skills
make list # Show available skills and install statusMost of these skills are designed to run as one loop, not as isolated commands. Set a project up once, capture work as it arrives, hand it to agents, check in, audit, and feed the audit back into the queue.
/project-setup (once per project)
│
▼
ideas, bugs ──► /todo ──► TODO.md ◄──────────────────┐
your own asks ──────────────► ASKS.md │
│ │
▼ │
/bug-bash one agent, while you watch │
/go-team a crew, unattended — overnight │
│ │
▼ │
/sitrep what landed · what's running │
· what's blocked on YOU │
│ │
▼ │
/scorecard ── findings ────────────────────┘
1. Set up once — /project-setup. Walks through the scaffolding the rest of the loop depends on: a tracker (TODO.md), an engineering diary, CLAUDE.md rules, and — the one that decides whether an overnight run survives — a permissions allowlist so agents don't stall on a prompt at 2am. Run it again later with status to see what's missing.
2. Capture work as it comes — /todo. Braindump bugs and ideas into the tracker as you notice them, without breaking flow. Your own requests go in ASKS.md, a separate lane that agents rank above everything they found for themselves — otherwise a feature you designed sits unbuilt for twelve hours while agents fix bugs in the thing it was meant to replace.
3. Do the work. Two modes:
/bug-bash— one agent works the tracker in priority order while you're around./go-team— a crew of three seats runs unattended: a lane seat on your own asks (and other projects' requests), a product seat on the roadmap, and a consolidation seat whose only job is removal — the counter-force a fleet otherwise lacks. Built on one premise: agents systematically overstate what they've done, so nothing merges on a report — every claim is reproduced in an isolated worktree, the gate reads a verdict file rather than the worker, and it checks shape (thesis, surface area, fix-by-addition) as well as correctness. A compass (/slopand/scorecardevery ten cycles) redirects effort when the codebase is getting worse, and a stop condition pauses a fleet that's only feeding itself. It degrades rather than stalls. Run it on/loopand go to bed. Either way, commit often and run/consolidatebefore pushing — checkpoints are for you, the history is for everyone after — and a machine-wide pre-push hook refuses a push that skipped it.
4. Check in — /sitrep. When you come back: what landed, what's still running, and — first, above everything — anything waiting on a decision or permission only you can give. A blocked agent is the only line in a status report still costing you something while you read it.
5. Audit, and feed it back — /scorecard. Periodic letter-graded health check across 13 code dimensions plus a second table for agent-readiness: could an agent work here for eight hours without a human? Its Security grade cross-checks against /delegate-security-audit for a second model's opinion. Findings go back into the tracker, where /go-team picks them up.
Across projects — /request. Projects depend on each other, and the old habit of jumping into ../otherproject to fix it yourself put agents in each other's checkouts. Now you file a request into that project's inbox (a file outside every repo); their agents triage it after your own asks and before their self-generated backlog, and /sitrep shows both what's been asked of you and what you're waiting on. A hook refuses edits to foreign repositories, so the rule has a mechanism.
5b. Watch for slop — /slop. Cheap code has a failure mode that looks fine commit by commit: every change adds a guard, a fallback, a special case, and nothing is ever folded back. /slop measures whether a repo is accreting — splitting history into a hand-written era and an AI era when it can — and, run on a branch, asks the one question that separates a chosen change from an unchosen one: could the author say why? It grades on articulability and surface area, explicitly not on diff size; a big refactor with one thesis passes, a small fix with none doesn't.
| Skill | Role in the loop |
|---|---|
project-setup |
Establish the scaffolding once (tracker, diary, rules, unattended permissions) |
todo / bug |
Capture work into the tracker without breaking flow |
bug-bash |
Work the tracker with one agent, attended |
go-team |
Work the tracker with a crew, unattended; verify before merging |
sitrep |
Resume: what landed, what's running, what's blocked on you |
scorecard |
Periodic audit; findings feed back into the tracker |
slop |
Audit for unchosen code — accretion trends on a repo, or a pre-merge check on a branch |
request |
Ask another project's agents to do work — a mailbox outside every checkout, so nobody edits anyone else's repo |
consolidate |
Rebuild unpushed commits into one per feature or fix, tree-identical; a machine-wide pre-push hook refuses pushes that aren't reader-ready |
retro |
Evidence-first retrospective from a session's transcripts and repo, when the process isn't working |
fleets |
Which go-team fleets are alive across every project — leases, last landing, session state |
For a specific task, outside the loop.
Delegating to other models. Claude can't generate images, and sometimes a second opinion from a different model lineage, a local-only model for private data, or a cheap model for grunt work is the right tool. These route through pi and OpenRouter, honouring a zero-data-retention-only policy.
| Skill | Description |
|---|---|
delegate-bulk |
High-volume, low-intelligence grunt work on a near-free model |
delegate-image |
Generate an image via a panel of top image models, with two independent AI judges picking their favourites |
delegate-private |
Work on private data with a local on-device model that never leaves the machine |
delegate-review |
Independent second-opinion code review from a different model lineage |
delegate-security-audit |
Deep security analysis and fixes via GLM |
Building and shipping.
| Skill | Description |
|---|---|
docs |
Plan and write user documentation — Diátaxis structure, audience first, a reader test before merge |
readme |
Generate or update project README documentation |
release-setup |
Set up automated cross-platform binary releases for a Go project |
tool-web |
Build a lightweight single-file web application with no external dependencies |
Test locally:
claude --plugin-dir .MIT