awesome-skills
English · 简体中文
A curated collection of the Agent Skills and workflows I've honed in real projects — encoding my engineering conventions and creative pipelines into capabilities that agents like Claude Code can call directly.
These aren't generic "do-anything assistants." Each skill takes a stance: my tech-stack conventions, my aesthetic bar, my collaboration patterns. Just describe what you need — the skill handles the architecture decisions, boilerplate, conventions, and delivery. Fewer detours, no cutting corners.
Tech-stack scaffolding
| skill | what it does |
|---|
| react-best-practices | React conventions & scaffolding: yarn + Vite + TS + Tailwind + shadcn/ui, with layering / naming / lint suite — init, build and review end to end |
| fastapi-best-practices | FastAPI backend conventions: uv + SQLAlchemy + Alembic + loguru + ruff, MVC layering, unified response envelope, GET+POST convention |
| docker-best-practices | Containerization standard kit: three-zone layout (images / containers / scripts), multi-image split, multi-arch buildx, registry & offline-tar delivery |
| website-creator | Spin up a website / app from scratch: Socratic questioning to 95% requirement certainty, then scaffolds frontend / full-stack |
Docs & knowledge
| skill | what it does |
|---|
| wiki-creator | Deep-scan a project into DeepWiki-style multi-file docs (Mermaid, Docsify-ready) |
| docs-site-creator | Create documentation sites with a two-provider router: Docsify (zero-build, offline single-file) for simple docs, Docusaurus 3 (module tabs / local search / dark mode / brandable theme) for professional portals |
| req-to-ai-spec | Turn scattered requirements (text / screenshots / legacy code) into structured, AI-friendly specs a coding agent can run with |
| spechub-best-practices | Write high-quality handoff specs managed via git worktree — built for AI-to-AI task handoff |
| notion-chat-archiver | Summarize an AI conversation and archive it into a personal Notion database (topic / takeaways / tags) |
| solution-vault | A personal solution library to replicate proven implementations (OAuth, uploads, payments…) across projects |
| pdf-craft | Typeset content into print-grade PDFs via HTML+CSS — four layout systems (government filing / business proposal / technical report / resume) plus a typography linter for Chinese print conventions |
Visualization
| skill | what it does |
|---|
| html-diagram | Pick a "face" from a style library and render architecture / topology / timeline diagrams as a self-contained single HTML file |
Design & style
| skill | what it does |
|---|
| style-vault | A six-tier (product / style / page / block / component / token) personal style library that generates frontends to my aesthetic |
| style-vault-sediment | The companion writer: deposit new styles into style-vault, versioned by author |
| slop-audit | Pre-ship "AI slop" audit: sweep a page against a countable checklist and a dictionary of model tells (decorative eyebrows, fake screenshots, locale strips, scroll cues, blacklisted palettes and serifs, AI-voice copy), each with a concrete replacement |
Image & video
| skill | what it does |
|---|
| comfyui | Drive a local ComfyUI from natural language for image / video generation, auto-building workflows; model-agnostic, tuned for Apple Silicon |
| codex-image-gen | Generate / edit images via Codex (gpt-image-2), with reference images to lock character & style |
| gemini-gen | Generate images via Gemini member accounts (Nano Banana) over browser cookies — multi-account load-balancing with auto-failover on quota |
| browser-gen | Drive a logged-in Gemini web session for free image / video generation (Veo video · Nano Banana image) |
| jimeng | Jimeng/Dreamina image + video via the cloud API (AK/SK, no vip needed): text/image-to-image, P-image edit, inpaint, text/image-to-video, and motion-mimic — drive a character image with a template video (IP animation made easy); CLI fallback |
| flow-gen | Generate videos and images in Google Flow with the credits of a Google AI Pro/Ultra plan: Veo 3.1 Lite/Fast/Quality & Omni Flash video from text, start/end frames or reference images (ingredients), plus free Nano Banana images (text or reference-to-image); drives your real logged-in Chrome via a pluggable bridge (default kimi-webbridge), no extra browser and no API key, full-quality mp4 with audio; models and costs read live from Flow |
| media-gen | A unified image-gen entry point: one call routes to gemini-gen / comfyui / codex / … by config & availability, with a quota-failover fallback chain, auto-injects style-lock v1, and returns a unified result contract (à la Pixelle's MediaService) |
Audio
| skill | what it does |
|---|
| voxcpm | Local text-to-speech / voiceover (VoxCPM2 on Apple MLX) — zero-shot, voice design from a text prompt, and voice cloning; offline, 48kHz, faster-than-realtime on Apple Silicon |
| sound-gen | Local text-to-music (BGM) + text-to-SFX generation — one prompt to instrumental background music (ACE-Step 1.5, MIT) or short sound effects (Stable Audio Open Small); offline, commercial-use, on Apple Silicon |
Methodology
| skill | what it does |
|---|
| skill-management | My skill-management methodology itself: organize many skills as source → category → skill, driven by a single registry, reusable across Claude Code, Gemini, Pi & Codex with per-agent sync whitelists and bundled tooling. Any AI can read it and replicate the whole setup |
| self-improving-workflow | A universal methodology: four reviewer sub-agents for collaborative learning + a single /run entrypoint for long-running autonomous execution |
| memory-palace | Guide anyone to build a platform-agnostic personal memory system: one Obsidian vault (plain Markdown) that Claude / Codex / Gemini read & write together, so every agent understands you over time — memory visible, editable, portable. Subcommands help/init/interview/extract/distill/review/analyze; deterministic zero-dep engine mp.py. Distils from Karpathy / OpenClaw / Hermes / open-second-brain / mem0 |
| loop-harness | Scaffold + harness for high-quality long-running autonomous iteration loops (optimization / audit / migration / deep grinding). Taskbook + ledger + log file-driven memory, an external objective verification gate (report what the script says, not self-judgment), progress-detection stop/escalate (not step count), anti-drift memory + multi-agent. Distilled from a real 100-round loop; isomorphic to Ralph loop + BabyAGI + evaluator-optimizer. Runs on top of /loop or /schedule |
| codex-conductor | "Claude as brain, Codex as hands" delegation workflow: dispatch implementation work to the local Codex CLI via the codex:codex-rescue agent — Claude decomposes, writes the task brief, independently verifies (Codex's own completion report never counts) and merges itself. Role presets (builder / coder / chore / detective / scout / reviewer — each a model-tier + effort + read-write + prompt-stance combo, one reference file per role), task-brief patterns, worktree discipline, background-job polling. Distilled from a real 9-work-package campaign fully implemented by Codex |
| rsi | Evaluation-driven RSI (recursive self-improvement) loop: say "start RSI for " and an orchestrator session runs rounds of scoring → a fresh subagent as improver (never sees the exam set) → a gate on frozen evaluation → merge, deploy, notify, until told to stop. Generic method; a project joins by adding one adapter file (commands, protected / frozen paths, cadence) |
Usage
These skills follow the Claude Code Agent Skills spec.
git clone https://github.com/garveyhu/awesome-skills.git
# Symlink the skills you want into Claude Code's skill dir (pick as needed)
ln -s "$PWD/awesome-skills/stack/react-best-practices" ~/.claude/skills/
ln -s "$PWD/awesome-skills/media/image/comfyui" ~/.claude/skills/
Then trigger them in conversation with natural language, or invoke /<skill-name> explicitly.
Shared DNA
- Single responsibility / extensible / maintainable: rather split one more layer than cram multiple responsibilities into one file.
- Anti "AI-slop" aesthetics: pick one "face" first, freeze the design tokens, build a single signature moment, keep the rest restrained.
- Convention first: backend GET + POST with a unified response envelope; frontend atomic styling, strict types.
- Chinese-first output: comments, docs and generated content default to Chinese.
A curated, opinionated collection — only my original skills. Use it, learn from it, and feel free to open an issue.
