lich-skills

Personal skill collection for Claude Code, Gemini CLI, and OpenAI Codex — by @LichAmnesia.

Telegraph-style, opinionated, no filler. Engineering judgment skills plus high-leverage utility tools.

中文版：README-zh.md

The go-no-go gate (Stage 0)

                          ┌─────────────────────────────────────┐
                          │   NO-GO is the default.             │
                          │   GO requires evidence.             │
                          └──────────────┬──────────────────────┘
                                         │
   MEMORY     →   HYPOTHESIS   →   FIVE CHECKS    →   24h INTERRUPT?   →   COMMITMENT
 ┌────────┐    ┌─────────────┐  ┌────────────────┐  ┌──────────────┐   ┌──────────────┐
 │ Search │    │ One-sentence│  │ Differentiation│  │ Enthusiasm   │   │ Public doc + │
 │ prior  │───▶│ guess +     │─▶│ Audience-fit   │─▶│ high? Wait   │──▶│ kill criteria│
 │ tries  │    │ confidence% │  │ Acquisition    │  │ 24h, restart │   │ D14/30/60/90 │
 └────────┘    └─────────────┘  │ Capacity       │  └──────────────┘   └──────┬───────┘
                                │ 7-Factor Wedge │                            │
                                └────────────────┘                            │
                                                                              ▼
                                                                       GO  → spec-driven-dev
                                                                       NO-GO → journal entry

Stage 0 before any /spec. Five framework gates, mandatory memory check against prior attempts, 24h pattern-interrupt for high-enthusiasm signals, public commitment artifact with pre-mortemed kill criteria. See skills/go-no-go/.

The spec-driven-dev loop

   DEFINE         PLAN           BUILD          TEST          REVIEW          SHIP
 ┌────────┐   ┌────────┐   ┌─────────┐   ┌────────┐   ┌────────┐   ┌────────┐
 │  Spec  │──▶│  Plan  │──▶│  Build  │──▶│  Test  │──▶│ Review │──▶│  Ship  │
 │  PRD   │   │ Tasks  │   │  Impl   │   │ Verify │   │  Gate  │   │  Tag   │
 └────────┘   └────────┘   └─────────┘   └────────┘   └────────┘   └────────┘
     ▲                                                                 │
     └─────────────────── feedback / regression ──────────────────────┘

One skill, six phases, explicit exit criteria per step. Pairs with go-no-go as Stage 1 of the pipeline. See skills/spec-driven-dev/.

The debug-hypothesis loop

                          ┌─────────────────────────────┐
                          │    all hypotheses rejected?  │
                          │    back to OBSERVE           │
                          └──────────┬──────────────────┘
                                     │
  OBSERVE  ──▶  HYPOTHESIZE  ──▶  EXPERIMENT  ──▶  CONCLUDE
     │              │                  │               │
     ▼              ▼                  ▼               ▼
  Reproduce      List 3-5          One test,       Root cause
  the bug,       causes +          max 5 lines,    confirmed
  collect        evidence           falsify         → fix +
  symptoms       for each          don't confirm   regression test
     │              │                  │               │
     │              │           ┌──────┘               │
     │              │           │ rejected?            │
     │              │           ▼                      │
     │              │     next hypothesis              │
     │              │                                  │
     └──────────────┴──── everything in DEBUG.md ──────┘

Scientific-method debugging. Prevents the #1 AI failure: bulldozing through a wrong theory. See skills/debug-hypothesis/.

The aggregate-N-sources loop

                    Trajectories-as-environment
            ╔════════════════════════════════════════╗
            ║   src_1   src_2   src_3  ...   src_N   ║
            ║   [..]    [..]    [..]         [..]    ║
            ║                                        ║
            ║   not concatenated. not summarized.    ║
            ║              navigated.                ║
            ╚═══════════════════╤════════════════════╝
                                ▼
            ┌────────────────────────────────────────┐
            │       AGGREGATOR  (lite agent)         │
            │   inspect_file / inspect_section       │
            │   search_sources / cross_pack_check    │
            │                                        │
            │   notes = []  (every claim → path:line)│
            │   budget = 25 → loop until coverage    │
            └───────────────────┬────────────────────┘
                                ▼
                  pack/  brief.md / findings.md /
                         sources.tsv / _aggregation_log.md

Agentic aggregation for long-horizon research. N raw notes → 1 structured pack with full path:line provenance + cross-source contradictions surfaced. See skills/wiki-aggregate/.

The build-until-pass loop

        ┌──────────────────────────────────────────────┐
        │  attempt > MAX (default 10)?                  │
        │  → STOP, report what's left, hand to human    │
        └───────────────────────┬──────────────────────┘
                                │ no
   RUN CHECK ──▶ exit 0? ──yes──▶  DONE  (green = exit code, not opinion)
       ▲             │
       │             │ no (red)
       │             ▼
       │      READ first error  ──▶  FIX smallest  ──▶  re-RUN
       │      verbatim, file:line    minimal diff,        │
       │                             never fake green     │
       └─────────────────────────────────────────────────┘
              one round = one fix + one re-run + one-line report

Turns the human out of the run → read error → fix one line → run again loop, and caps the agent so it can't bulldoze 200-line speculative rewrites between runs. Refuses to fake green (@ts-ignore, deleting tests, || true). See skills/build-until-pass/.

Skills

All skills read credentials from environment variables (TAVILY_API_KEY, GEMINI_API_KEY, etc.) — never hardcoded.

Quick install

/plugin marketplace add LichAmnesia/lich-skills
/plugin install lich-skills@lich-skills

/skills

# 1. Install Claude Code (if not already)
curl -fsSL https://claude.ai/install.sh | bash
# or: brew install --cask claude-code

# 2. Clone into the global skills directory
git clone https://github.com/LichAmnesia/lich-skills.git ~/.claude/skills/lich-skills

# 3. Start Claude Code
claude
> /skills

gemini extensions install https://github.com/LichAmnesia/lich-skills

gemini extensions list

npm install -g @google/gemini-cli
git clone https://github.com/LichAmnesia/lich-skills.git ~/.gemini/extensions/lich-skills

# 1. Install Codex CLI
npm install -g @openai/codex
# or: brew install --cask codex

# 2. Install the skill collection
mkdir -p ~/.codex/skills
git clone https://github.com/LichAmnesia/lich-skills.git ~/.codex/skills/lich-skills

Why the one-liners work

• /plugin marketplace add LichAmnesia/lich-skills — Claude Code reads .claude-plugin/marketplace.json at the repo root. That file declares the repo as a plugin marketplace and points the plugin back at its GitHub source.
• /plugin install lich-skills@lich-skills — format is <plugin-name>@<marketplace-name>. Both resolve from the same marketplace manifest, which is why the name appears twice.
• gemini extensions install <github-url> — Gemini CLI's native extensions subcommand clones any public GitHub repo that has a gemini-extension.json at its root, then auto-discovers bundled skills from skills/*/SKILL.md. The manifest is what makes this one-liner work — without it, Gemini CLI refuses to install the repo as an extension.
• geminicli.com/extensions/ listing — the public extension gallery sources third-party extensions from GitHub repos that have the same gemini-extension.json manifest. Having the manifest is necessary (though not sufficient) to appear there.
• git clone into ~/.claude/skills/ — the lowest-common-denominator path. Claude Code reads every SKILL.md under ~/.claude/skills/** on session start. No marketplace required.

Documentation

Security

No secrets committed. Repo scanned with gitleaks via pre-commit hook and CI on every PR. Skills use environment variables only; example configs use YOUR_API_KEY_HERE placeholders. Report any leaked credential via a private security advisory.

Contributing

PRs welcome. New skills should be specific, verifiable, and minimal. See CONTRIBUTING.md.

License

MIT © 2026 Shen Huang (@LichAmnesia)

Star History

{
  "name": "lich-skills",
  "owner": {
    "name": "Shen Huang (LichAmnesia)",
    "url": "https://github.com/LichAmnesia"
  },
  "metadata": {
    "description": "LichAmnesia's open-source Claude Code skills — skills-map router that picks which skill to use, go-no-go Stage 0 gate, spec-driven-dev workflow, debug-hypothesis, wiki-aggregate, Tavily search, Nano Banana image generation, frontend-design, and subagent-brief (pre-flight token discipline for multi-agent fan-out).",
    "version": "0.2.1",
    "license": "MIT"
  },
  "plugins": [
    {
      "name": "lich-skills",
      "source": {
        "source": "github",
        "repo": "LichAmnesia/lich-skills"
      },
      "description": "skills-map router (picks which skill to use) + go-no-go Stage 0 gate (NO-GO is the default) + Spec → Plan → Build → Test → Review → Ship workflow + debug-hypothesis + wiki-aggregate + Tavily web search + Nano Banana image generation + frontend-design aesthetics guide + subagent-brief token discipline + build-until-pass self-correcting build loop."
    }
  ]
}