Reusable agent skills and tools — original work, generalized from months of real-world use.
Skills follow the open Agent Skills standard (SKILL.md + YAML frontmatter) and work with Claude Code, OpenAI Codex, pi, Cursor, GitHub Copilot, Gemini CLI, opencode, Goose, Amp, and any other compatible harness.
npx skills add iksnae/skills # universal installer — auto-detects your agents/plugin marketplace add iksnae/skills # Claude Code
/plugin install iksnae-skills@iksnae
npx @iksnae/skills list # this repo's own installer
npx @iksnae/skills add certify chaos-qa # → ~/.claude/skills + ~/.agents/skills
npx @iksnae/skills add --all --project # → ./.claude + ./.agentsManual: copy any skills/<name>/ into ~/.agents/skills/ (Codex, pi, Cursor, opencode, Goose, Amp) or ~/.claude/skills/ (Claude Code).
Every skill has a full page covering usage, behavior, outputs — and a real demo run against nightjar, the fictional pastebin built to be inspected, broken, and fixed by these skills.
| Skill | What it does |
|---|---|
| certify | Score every code unit across 9 quality dimensions; graded report card plus remediation plan. |
| repo-audit | Read-only health and due-diligence audit of an unfamiliar repository. |
| state-of-project | Live "state of the project" board — shipped vs. in-progress vs. planned, in the project's own brand and status vocabulary, deployed as an Artifact. |
| market-scout | Comparative research with adversarial fact-checking and a weighted, cited scorecard. |
| retrospective | Evidence-only retrospectives — every claim cites a commit, run, or issue. |
| Skill | What it does |
|---|---|
| development-loop | Language-agnostic plan → implement → review → refactor loop with clean-code discipline. |
| tree-shaking | Shrink binaries and bundles — measure, diagnose, pull the right dead-code/packaging lever, prove the win. Language specialties (Swift, Go, JS/TS) under one skill. |
| docctor | Author Apple DocC documentation — symbol comments, .docc catalogs, Topics curation, tutorials, and static hosting to GitHub Pages. |
| Skill | What it does |
|---|---|
| grumpy | Skeptical senior review — finds where code, plans, or APIs will hurt later; severity-ranked smells with fixes and named test gaps. |
| janitor | Disciplined, behavior-preserving cleanup — the smallest safe patch that makes the next change less scary; pairs with grumpy. |
| Skill | What it does |
|---|---|
| chaos-qa | Adversarial chaos GameDays: inject faults, observe, file findings with repro. |
| dogfood-qa | Exercise an app end-to-end as a real user following its own docs. |
| surface-consistency-audit | Cross-check the same fact across CLI, API, and web surfaces; classify the drift. |
| Skill | What it does |
|---|---|
| image-generate | Brand-aware image generation with receipts — this README's heroes are its output. |
| illustrate-doc | Decide whether a document deserves illustration, then orchestrate the media skills. |
| article-audio | Narrated audio from a markdown article, with pronunciation config. |
| spoken-updates | Toggleable, configurable spoken agent progress updates via non-blocking native TTS. |
| remotion-author | Author Remotion video specs in a lintable blueprint format. |
| remotion-render | Render Remotion compositions with receipts; bundles the spec linter. |
| remotion-with-image | Composite flow: generated stills inside Remotion videos. |
| Skill | What it does |
|---|---|
| familiar | A cross-harness ambient desktop pet that reacts to your agent's activity — semantic events → a pure reducer → renderer-agnostic pets (macOS overlay or terminal). |
| pet-hatch | Author the looping sprite animations a familiar pet plays — one loop per semantic state, generated as planted strips, sliced to frames + an anim.json. |
Claude Code slash commands bundled in .claude/commands/:
| Command | What it does |
|---|---|
/ask-more |
Prompts proactive clarification via the AskUserQuestion tool whenever something is less than ~85% clear, with auto-answer defaults that prefer the correct, best-practice solution over the quick fix. |
nightjar is a small Go terminal pastebin that exists to be demonstrated on. The skills in this repo audited it, chaos-tested it (finding ~19% silent write loss under concurrency), dogfooded it (hitting a documented-but-missing nj rm), fixed it through the development loop, picked its v2 storage engine, and produced its launch imagery, narration, and title-card spec. Read the whole arc.
skills/<name>/SKILL.md # one directory per skill (Agent Skills standard)
docs/<name>.md # full page per skill, with demo
docs/demos/ # real artifacts from the demo runs
demo/nightjar/ # the fictional project the demos run against
.claude/commands/ # bundled Claude Code slash commands
.claude-plugin/ # Claude Code marketplace + plugin manifests
tools/ # runnable companions (e.g. the familiar renderer)
bin/cli.mjs # npx installer
- Skills are written to be model-invoked by their
description— no harness-specific invocation required. - Claude-specific frontmatter (
allowed-tools, slash-command hints) is advisory; other harnesses ignore it safely. market-scoutbundles an optional Claude Code Workflow script; the skill works without it everywhere else.- Media skills shell out to bundled Python scripts in each skill's
scripts/dir and read credentials from environment variables (OPENAI_API_KEY) — harness-agnostic.
MIT © iksnae