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 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. 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/.

---

Skills

Skill	What it does
spec-driven-dev	Full SDLC workflow: Spec → Plan → Build → Test → Review → Ship. Anti-rationalization tables, verification gates, atomic commits.
debug-hypothesis	Scientific-method debugging: Observe → Hypothesize → Experiment → Conclude. Anti-bulldozer rules, max 5-line experiments, mandatory DEBUG.md evidence trail.
wiki-aggregate	Lift N≥3 raw research artifacts into one structured pack via agentic aggregation. Cheap-pass + tool-budgeted aggregator loop, every claim has path:line provenance, cross-source contradictions logged.
tavily-search	Web search + content extraction via the Tavily API. Use for fact-checking, docs lookup, source-cited research.
nano-banana	Text-to-image and image editing via Google's Nano Banana 2 (gemini-3.1-flash-image-preview). Supports 512 / 1K / 2K / 4K.
frontend-design	Build distinctive, production-grade frontend interfaces — bold aesthetic direction, intentional typography, and motion that avoids generic AI-slop UI. Adapted from Anthropic's official frontend-design skill (Apache-2.0).
subagent-brief	Pre-flight discipline for spawning subagents. Anthropic does NOT share prefix across subagents — each one cold-starts on the full prompt. Compress every subagent prompt into a ≤200-word brief before spawning. Five rules + brief template + anti-rationalization table. Backed by arXiv 2604.25899 (Pythia, 2026).

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

---

Quick install

Claude Code — plugin marketplace (one command)

Inside a running Claude Code session:

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

Done. All seven skills become available immediately. Verify:

/skills

Claude Code — git clone

# 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

Full guide: docs/claude-code-setup.md.

Gemini CLI — extensions install (one command)

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

This repo ships a gemini-extension.json at the root, so Gemini CLI installs it as a first-class extension and auto-discovers every skills/*/SKILL.md. Verify:

gemini extensions list

Manual clone fallback:

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

Full guide: docs/gemini-cli-setup.md.

OpenAI Codex CLI

# 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

Full guide: docs/codex-setup.md.

---

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

Tool	Install + skill setup
Claude Code	docs/claude-code-setup.md
Gemini CLI	docs/gemini-cli-setup.md
OpenAI Codex	docs/codex-setup.md

---

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