luci-efe/skills
Agentic engineering skills by Fernando Ramos. Agents are not chat windows: they should be observable, talk to each other directly, prove they're done, carry context forward, mix models, and act with bounded ownership.
Español abajo · Site: ferramos.com/skills
Install
npx skills@latest add luci-efe/skills # all skills, any supported agent
npx skills@latest add luci-efe/skills -s proof -s relay # pick skills
Claude Code plugin:
claude plugin marketplace add luci-efe/skills
claude plugin install ramos-skills@ramos
Then run /doctor in your repo. It lists missing tools with the exact install command, and can wire the optional Claude Code hooks.
Works in your agent
The skills follow the open Agent Skills format (SKILL.md with name and description), so one install works across harnesses. npx skills writes to each agent's own folder:
| Agent | Install | Lands in |
|---|
| Claude Code | npx skills@latest add luci-efe/skills -a claude-code, or the plugin below | .claude/skills/ |
| Codex, Gemini CLI, Cursor, GitHub Copilot, OpenCode, Antigravity | npx skills@latest add luci-efe/skills -a codex (swap the agent id) | .agents/skills/ |
Any other of the 70+ agents skills supports | npx skills@latest add luci-efe/skills and pick yours | its own skills folder |
Add -g for a global install and --copy to copy instead of symlink. Browse them on skills.sh.
What differs by agent: the hard /proof Stop gate works in Claude Code and Codex (watched live on Codex) and is wired for Gemini CLI from its docs (not watched live); the other hooks are Claude Code only. Elsewhere the same scripts run as manual checks (see /doctor). /council and /relay launch need Herdr.
The skills
| Stage | Skill | Use when | Rule |
|---|
| Orient | /doctor | Setting up, or a skill reports a missing tool | 06 |
| Orient | /compass | You aren't sure which skill fits. It routes, and checks tools first | 06 |
| Orient | /cartographer | You need one domain or flow mapped to real code | 04 · 01 |
| Decide | /crucible | A change is underspecified. It reads code first and asks only the open questions | 06 |
| Decide | /council | A big decision deserves review by several model families in Herdr | 02 · 05 |
| Build | /reproduce | Something is broken. It turns the bug into a failing test first | 03 · 01 |
| Prove | /proof | You're about to say "done". It records checks against the exact code state | 03 · 01 |
| Prove | /red-team | Dependencies changed or an advisory landed | 03 |
| Continue | /relay | A session hands work to another session, agent, or machine | 04 · 02 |
Rules: 01 Observable · 02 Not a clipboard · 03 Evidence of done · 04 Context continues · 05 Many models, many harnesses · 06 Autonomy with ownership.
How it behaves
- Fails fast on missing tools. Each skill branch declares its tools in
skill.json. A missing tool blocks only that branch and prints the install command. Nothing silently degrades.
- One hard gate. With hooks installed, once a Claude Code session has a passing
/proof run (which arms the gate for that session), the session can't end while its evidence is missing, failing, or stale for the current code, unless the run is recorded as blocked, failed, or cancelled with a reason. The gate doesn't read the conversation: a session without a passing /proof run isn't blocked. Every other hook is advisory. In other harnesses the same scripts run as manual checks.
- State lives in
.ramos/ (override with RAMOS_STATE_DIR). Add it to .gitignore unless you want evidence committed.
- "Done" vs. "verified". The local gate prevents accidental false completion claims; hand-written evidence can still pass it. CI re-execution plus independent review earns
verified.
Tools
| Tool | Needed by | Install (macOS) |
|---|
| node ≥22, git | all scripts | brew install node git |
| ripgrep | cartographer, crucible, red-team | brew install ripgrep |
| ast-grep | cartographer, red-team | brew install ast-grep |
| mnemosyne | relay (required) | uv tool install --python 3.12 mnemosyne-memory (needs uv: brew install uv) |
| herdr | council, relay launch | brew install herdr (herdr.dev) |
| osv-scanner | red-team | brew install osv-scanner |
| gitnexus | cartographer (optional) | npm i -g gitnexus. License: PolyForm Noncommercial. Check it before commercial use. |
| playwright | proof visual-web | per repo: npm i -D @playwright/test then npx playwright install chromium |
| odiff (optional) | proof visual-web pixel diff | per repo: npm i -D odiff-bin (or use Playwright's toHaveScreenshot) |
| cua-driver | proof visual-desktop | from cua.ai (macOS). This suite doesn't verify the install; /doctor only checks that cua-driver is on PATH |
Develop
Requires node ≥22, git, ripgrep, ast-grep, and mnemosyne on PATH (the relay memory tests and the cartographer ast-grep test must not skip).
npm test # node --test
npm run lint:skills # frontmatter, size, skill.json schema
npm run check:vendor # shared lib copies match lib/
Docs: per-skill pages and memory · Worked examples · Contracts: docs/CONTRACTS.md.
Español
Skills de ingeniería agéntica por Fernando Ramos. Los agentes no son ventanas de chat: deben ser observables, hablar entre sí directamente, demostrar que terminaron, conservar el contexto, combinar modelos y actuar con autonomía acotada.
Instalación
npx skills@latest add luci-efe/skills
npx skills@latest add luci-efe/skills -s proof -s relay
Plugin de Claude Code: claude plugin marketplace add luci-efe/skills y luego claude plugin install ramos-skills@ramos. Después ejecuta /doctor en tu repositorio.
Cómo se comporta
- Falla rápido si falta una herramienta. Muestra el comando exacto de instalación y nunca degrada en silencio.
- Una sola compuerta obligatoria. Con los hooks instalados, una vez que una sesión de Claude Code tiene una ejecución de
/proof que pasó (eso activa la compuerta para esa sesión), la sesión no puede terminar si su evidencia falta, falló o quedó desactualizada respecto al código, salvo que la ejecución quede registrada como bloqueada, fallida o cancelada con un motivo. La compuerta no lee la conversación: una sesión sin una ejecución de /proof que pasó no se bloquea. Los demás hooks solo sugieren.
- El estado vive en
.ramos/ (se puede cambiar con RAMOS_STATE_DIR).
- "Terminado" vs. "verificado". La compuerta local evita declaraciones accidentales de finalización; una evidencia escrita a mano aún puede pasarla. La reejecución en CI y una revisión independiente permiten llamarlo verificado.
Los comandos y nombres de archivo se mantienen en inglés.
- Funciona en tu agente. Sigue el formato abierto Agent Skills. Claude Code:
npx skills@latest add luci-efe/skills -a claude-code. Codex, Gemini CLI, Cursor, GitHub Copilot, OpenCode y Antigravity: -a codex (cambia el id del agente). La compuerta de /proof funciona en Claude Code y Codex (vista en vivo en Codex) y está conectada para Gemini CLI según su documentación (no la vimos en vivo); los demás hooks son solo de Claude Code. En el resto de agentes los mismos scripts se ejecutan manualmente.
License
MIT © Fernando Ramos
