skills

Agent skills for long-running engineering work. Written once as SKILL.md, usable in every agent that reads the open standard — Claude Code, Codex, Gemini CLI, Cursor and others.

No conversion step and no per-tool copies. A skill here is a folder with a SKILL.md in it, which is exactly what each of those tools already looks for. The installer's whole job is putting the folder where each one looks.

The skills

Skill	What it does
relay	Unattended autonomous work for a stated duration — 4 hours or 24 — on its own worktree and draft PR. Each iteration is a fresh process handing a committed ledger to the next, so nothing lives long enough to suffer context rot.
shred-plan	The adversarial pass between "here's the plan" and "let's build it". Verifies every codebase claim in a plan with evidence, then hunts hidden dependencies, missing failure modes, scope leaks and unverifiable done-criteria — findings ranked by severity, no mercy.
draft-plan	Turn an investigation or conversation into a written plan doc: verified diagnosis, explicit in/out scope, workstreams with proofs of done, sequencing gates, open questions. The doc shape shred-plan was built to check — chain them.
reclaim	Give a pinned dev machine back. Reads the System-vs-User CPU split first, because the process at the top of the list is usually a symptom — launchservicesd at 300% is churn, not work. Reaps stale build watchers, caps the test runner, and says plainly when only a reboot will do.
telegram-notify	Telegram bot pings when an agent finishes a task (Stop hook), lands a commit, or a relay run starts/lands/halts/ends. Ships the sender script and walks through BotFather setup and hook wiring.
safe-migration	The boring, safe Drizzle migration ritual. Reads the repo's own rule first — a frozen journal means hand-written, idempotent SQL and never the generator — takes the next prefix from disk and the default branch, greps for silent DROPs, and lands schema and migration in one commit.
vercel-triage	Turn a raw vercel logs dump into a triage: dedupe the CLI's repeated events, group by status and path, 5xx first. Read-only.
morning-brief	A founder's start of day in one screen: repos and PRs, an opt-in business block (inbox waiting on you, Stripe events, signups, prod errors, overdue tasks — from whatever tools are connected), news from your own feeds, and a Top 3 for today.
worktree-janitor	Eighty agent worktrees, one table: uncommitted, unpushed, PR state, the dev server holding a port, and a SAFE / ASK / KEEP verdict. Removes only what you pick, after re-checking each.
memory-hygiene	Claude Code loads only the head of MEMORY.md. Shows how much never reaches the agent — usually the rules — then archives finished project memories, shortens the index and moves rules to the top. Plus a session-wrap mode.

Install

One command, every agent you have. This uses the community skills CLI, which resolves the repo, symlinks each skill into every agent directory it detects, and records a lockfile so update works later:

npx skills@latest add ritmillio/skills

npx skills@latest add ritmillio/skills --list          # see what is here first
npx skills@latest add ritmillio/skills --skill relay   # just one
npx skills@latest add ritmillio/skills --all           # every skill, every agent
npx skills@latest list                                 # what you have installed
npx skills@latest update                               # pull newer versions

Claude Code can also install through the plugin marketplace, which gives you versioned updates via /plugin update:

/plugin marketplace add ritmillio/skills
/plugin install relay@ritmillio-tools

One catch, verified rather than assumed: a marketplace-installed skill is namespaced, so it invokes as /relay:relay. Skills installed by the CLI keep the plain /relay. The documentation implies a single-skill plugin escapes the namespace; on Claude Code 2.1.x it does not.

Companion packs

The marketplace also passes through Matt Pocock's skills (grill-me, tdd, to-spec, to-tickets, implement, code-review, …) as a remote-source entry — no vendored copy, nothing to drift:

/plugin install mattpocock-skills@ritmillio-tools   # pulled from his repo, synced by /plugin update

Skills CLI users get the same pack straight from the source instead:

npx skills@latest add mattpocock/skills

His pack is read-only either way. If you want to hack on one of his skills, copy it into skills/ here, keep the credit, and make it your own.

Layout

skills/
  engineering/     shipped, code-facing
  productivity/    shipped, everything else
  in-progress/     written but not trustworthy yet
  deprecated/      retired, kept because the reasoning is still worth reading

A skill is a folder with a SKILL.md, optionally beside references/, scripts/ and assets/. Nothing else is required, and the category is just a directory: installers find skills by scanning, not by path.

What "works everywhere" actually means

SKILL.md is an open format: YAML frontmatter with a name and a description, then Markdown instructions, optionally beside references/, scripts/ and assets/. Agents read the description to decide when a skill is relevant, then load the body.

That much is genuinely portable. Two things are not, and this repo is explicit about both rather than papering over them:

• Slash-command invocation is a Claude Code convenience. In other agents you ask for the skill by name or let the description trigger it.
• Bundled scripts need an absolute path, and the skill directory sits somewhere different in each tool. Skills here resolve it at runtime instead of assuming, and relay shows the pattern.

Add your own

./scripts/new-skill.sh my-skill --description "What it does and when to use it"

Scaffolds skills/my-skill/SKILL.md, writes the Claude Code plugin manifest, and registers it in the marketplace so it is installable rather than merely present. Then write the body and run ./scripts/install.sh my-skill.

See CONTRIBUTING.md for what makes a skill worth writing.

License

MIT