lich-skills
Battle-tested agent skills for Claude Code, Gemini CLI & Codex — engineering judgment as code.
An open-source collection of engineering-judgment and utility skills for Claude Code, Gemini CLI, and OpenAI Codex. Each skill packages a real workflow — research loops, critic passes, deploy runbooks — so your agent inherits judgment, not just tools.
What is lich-skills?
lich-skills is an open-source (MIT) collection of agent skills for Claude Code, Gemini CLI, and OpenAI Codex by @LichAmnesia — telegraph-style, opinionated engineering-judgment workflows plus utility tools. Its core is a set of disciplined loops: a go-no-go Stage 0 gate where NO-GO is the default, a six-phase spec-driven-dev SDLC, scientific-method debug-hypothesis debugging, wiki-aggregate research aggregation with path:line provenance, and a bounded build-until-pass fix loop — installable in one command via the Claude Code plugin marketplace or `gemini extensions install`.
Key features
- Skills for three agent CLIs: Claude Code, Gemini CLI, and OpenAI Codex, with per-tool setup docs
- go-no-go: a Stage 0 project gate where NO-GO is the default — memory check against prior attempts, 5 framework gates, a 24h pattern-interrupt for high enthusiasm, and a public commitment doc with D14/D30/D60/D90 kill criteria
- spec-driven-dev: full SDLC in six phases (Spec → Plan → Build → Test → Review → Ship) with explicit exit criteria, anti-rationalization tables, and verification gates
- debug-hypothesis: Observe → Hypothesize → Experiment → Conclude scientific-method debugging with max 5-line experiments and a mandatory DEBUG.md evidence trail
- build-until-pass: bounded self-correcting loop (run → read first error → smallest fix → re-run) capped at 10 attempts; green means exit code 0, and it refuses to fake green via @ts-ignore, deleted tests, or `|| true`
- wiki-aggregate: lifts N≥3 raw research artifacts into one structured pack where every claim carries path:line provenance and cross-source contradictions are logged
- skills-map router plus utility skills: tavily-search (web search), nano-banana (image gen/edit), frontend-design, google-analytics SEO reports, and subagent-brief
- One-command install: `/plugin marketplace add LichAmnesia/lich-skills` in Claude Code, or `gemini extensions install <repo URL>` (repo ships a gemini-extension.json manifest); git-clone fallback for Codex
- Security: no secrets committed — credentials via environment variables only, repo scanned with gitleaks in pre-commit and CI
Who it's for
- Claude Code / Gemini CLI / Codex users who want disciplined, verifiable workflows instead of ad-hoc prompting
- Solo founders and indie devs deciding whether to start a project at all — the go-no-go gate ships starter packs for solo-founder, indie-dev, and content-creator
- Engineers who want their agent to debug scientifically and fix builds in bounded loops rather than bulldozing speculative rewrites
- Researchers aggregating multiple raw notes into one provenance-tracked pack
- Anyone spawning subagents who wants the subagent-brief discipline (≤200-word briefs, since prefix isn't shared across subagents)
When not to use it
This is a personal, opinionated collection — telegraph-style with strong process opinions (e.g., NO-GO by default, hard attempt caps) — so it's a poor fit if you want a neutral, exhaustive skill library or work outside Claude Code, Gemini CLI, and Codex. Some utility skills also require third-party API keys (Tavily, Gemini) to function.
FAQ
How do I install lich-skills in Claude Code?
Two commands inside a session: `/plugin marketplace add LichAmnesia/lich-skills` then `/plugin install lich-skills@lich-skills`. All skills become available immediately; verify with `/skills`. A git-clone into ~/.claude/skills/ also works.
Does lich-skills work with Gemini CLI and OpenAI Codex too?
Yes. Gemini CLI: `gemini extensions install https://github.com/LichAmnesia/lich-skills` (the repo ships a gemini-extension.json so it installs as a first-class extension). Codex: git clone into ~/.codex/skills/. Setup guides live in docs/.
What is the go-no-go skill?
A Stage 0 gate run before any spec: NO-GO is the default and GO requires evidence. It checks memory of prior attempts, 5 framework gates, adds a 24h pattern-interrupt for high enthusiasm, and ends in a public commitment doc with D14/D30/D60/D90 kill criteria.
How does build-until-pass avoid faking a green build?
Green is the exit code, not the agent's opinion. Each round is one smallest fix + one re-run, capped at 10 attempts, and it explicitly refuses fake-green moves like @ts-ignore, deleting tests, or appending `|| true`.
Are API keys or secrets stored in the repo?
No. Skills read credentials from environment variables only (TAVILY_API_KEY, GEMINI_API_KEY, etc.), example configs use placeholders, and the repo is scanned with gitleaks via pre-commit hook and CI on every PR.
Share this launch
Embed this badge
<a href="https://orangebot.ai/product/lich-skills" target="_blank" rel="noopener noreferrer"> <img src="https://orangebot.ai/api/badge/lich-skills.svg" alt="Featured on OrangeBot" width="200" height="54" /> </a>