AX-skills: Skills Claude de Produção
xingfanxia/AX-skills
Skills reutilizáveis production-grade para Claude Code — marketing, escrita, pesquisa, mídia e mais.
Ver no GitHubColeção de skills production-grade para Claude Code, construídos por @xingfanxia. Inclui o skill jewelry-marketing (uma foto → pacote completo de marketing XHS com 12 imagens, 6 variações de copy e análise), além de skills para escrita, pesquisa, geração de mídia, e coaching de aprendizado. Cada skill é auto-contido, testado e pronto para produção.
Instalação
npx convez-ai xingfanxia/AX-skills1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Gerar pacotes de marketing completos para joias em e-commerce via uma única foto
- Criar conteúdo de escrita com estilos específicos (Khazix, LatePost)
- Conduzir pesquisa verificada multi-agente com rastreamento de fontes
- Gerar imagens com GPT-4 ou Gemini com fallback automático
- Transcrever áudio/vídeo com diarização de locutor
Veredito

Portfólio maduro de skills especializados por domínio com código aberto e documentação exemplar. O skill jewelry-marketing é um case de uso bem definido para e-commerce. Recomendado para quem constrói workflows verticalizados com Claude Code.
README
A tradução para o português deste README ainda está na fila.
AX-skills
A collection of production-grade Claude Code skills built by @xingfanxia (AX). Each skill is self-contained, runs on Claude Code / OpenAI Codex / OpenClaw, and ships with examples + tests.
What's a Skill? A reusable slash-command-style capability for AI coding agents — bundles a
SKILL.md(how to use), Python entrypoints, prompt templates, and reference docs. The agent invokes it like/jewelry-marketingand the skill takes over.
Skills in this repo
Vertical / domain skills
| Skill | Purpose | Status |
|---|---|---|
jewelry-marketing | One photo in → full XHS marketing bundle out (12 images + 6 copy styles + analysis) for jewelry e-commerce merchants. Auto-routes finished products vs raw stones. | ✅ v1.0 |
banxian-skill | 三合一东方占卜(小六壬 / 梅花易数 / 六爻)+ 赛博半仙人设。py 算法引擎从 panpanmao 玄学平台 TS 移植,64 卦完整数据 + 守"一事一占""医不问卦"规矩。 | ✅ v1.0 |
pet-forge | AI desk-pet asset generator — text or photo → 10 character-consistent APNG state animations + ready-to-install theme.json for clawd-on-desk (or any APNG-based pet runtime). Productionizes the 小肥 + 胖猫 build workflow. ~$2.30/pet, ~6-8 min. | ✅ v1.0 |
learning-coach | 在 Claude Code 里主动引导你学任何东西的学习教练(语言/技术/技能)。轻校准→喂 i-1 可理解输入→零评判猜读→工程化撞见→"今天懂了"→主动给下一小步。全程零考试·零背诵·零纠错羞辱,降焦虑优先于进度。领域无关内核 + 可选 playbook(英语首个)。四方调研交叉验证支撑。 | 🚧 v0.2 MVP |
Writing
| Skill | Purpose | Status |
|---|---|---|
khazix-writer | 卡兹克风格中文公众号长文 — 强口语化、叙事驱动、四层自检(L1 硬性规则 → L2 风格 → L3 内容 → L4 活人感)。Personal-experience / methodology long-form. | ✅ |
wandian-writer | 晚点 LatePost 风格中文深度长文 — 冷静判断、数据先行、判断式 header、段尾落点、对立面陈述。Industry analysis / company breakdown. Includes 6-范文 in-context reference. | ✅ |
Research
| Skill | Purpose | Status |
|---|---|---|
deep-research | Verification-style research — multi-agent parallel investigation, Tier 1-4 source authority cross-validation, traceable decision memo output. Use when judging claims to make a decision. | ✅ |
narrative-research | 横纵分析法(Horizontal-Vertical Analysis)by 卡兹克 — 双轴叙事型调研。纵轴追溯生命历程,横轴对比共时竞品,交叉出洞察 + PDF 报告。Use for understanding a thing from zero. Narrative counterpart to deep-research. | ✅ |
Media generation
| Skill | Purpose | Status |
|---|---|---|
gpt-image | OpenAI gpt-image-2 image generation — Azure-routed with auto fallback to OpenAI direct on rate-limit. Defaults to JPG. Best for photorealistic / editorial / product / UI mockups / images with rendered text. Sister to gemini-image. | ✅ |
gemini-image | Google Gemini 3.1 Flash Image generation. Sister to gpt-image — same shape, different provider. Best for multi-reference image input (background swap, character consistency across scenes — gpt-image cannot do this) and softer / illustrated / Studio Ghibli / watercolor aesthetics. | ✅ |
transcribe | Audio / video transcription via Google Gemini 3 Flash. Speaker diarization, auto language detection, files up to 500MB / ~8.4 hours. ~$0.50/M input tokens. | ✅ |
Documents
| Skill | Purpose | Status |
|---|---|---|
apple-pdf | Markdown / notes / reports → professionally formatted PDFs with SF typography. Default PDF generator. | ✅ |
Workflow / quality
| Skill | Purpose | Status |
|---|---|---|
mtc | More Than Coding — full product workflow from concept to code. Iterates design / story first → aligned data → build → polish. For games, demos, prototypes, interactive experiences. The signature AX workflow. | ✅ |
plan-design-review | Design-completeness review for PLAN.md / spec docs. Rates 7 dimensions 0-10, detects AI-slop patterns, builds interaction state tables. Catches gaps before code goes in. | ✅ |
codebase-sweep | Full-codebase audit + cleanup loop (parallel reviewers + iterative fix + docs cleanup + architecture documentation + orphan-script archiving). One-shot comprehensive review of the whole project. | ✅ |
audit-fix-loop | Multi-round code-quality pass scoped to recent changes. Loops until no high-confidence issues remain (max 5 rounds). Each round adds tests for fixed code + incremental doc updates. | ✅ |
pr-fix-loop | Autonomously respond to GitHub PR review feedback until all comments resolved + CI green. Polls all four comment surfaces — @claude posts to issue thread, not formal reviews, so single-surface polls miss it. Max 5 rounds. | ✅ |
Meta / knowledge
| Skill | Purpose | Status |
|---|---|---|
neat-freak | 洁癖 — End-of-session knowledge sync. Reconciles agent memory + project root markdown + docs/ + README against actual code so nothing rots. Three-audience editorial pass (agent / project-AI / external readers). Cross-platform (Claude Code · Codex · OpenCode · OpenClaw). Idempotent — safe to run every phase. | ✅ |
Install (Claude Code)
# Clone the repo
git clone https://github.com/xingfanxia/AX-skills.git ~/AX-skills
# Symlink the skill into Claude Code's skill dir
mkdir -p ~/.claude/skills
ln -sf ~/AX-skills/jewelry-marketing ~/.claude/skills/jewelry-marketing
# Reload Claude Code — the skill should now show up via /<skill-name> or auto-trigger
For OpenAI Codex / OpenClaw, follow your platform's skill installation convention (usually a similar symlink pattern).
Skill structure (convention)
<skill-name>/
├── SKILL.md # Required — frontmatter (name + description) + usage doc
├── README.md # GitHub-facing overview (optional, can mirror SKILL.md)
├── generate.py # Main CLI entry point (PEP 723 uv script preferred)
├── prompts/ # Prompt builders, helpers, output formatters
├── references/ # Domain-specific reference docs
└── examples/ # Test fixtures (small images, sample inputs)
Why these skills
- Distilled from real production: Each skill is extracted from a working product, not a toy demo.
jewelry-marketingcomes from shichuan (识川) — research-backed by 千瓜 / 数英 case studies on 周大福 / HEFANG / 樱桃珠宝 / 翡翠平安环 etc. - Niche depth over generic breadth: Better to nail one merchant vertical than be average at everything.
- Tool-grade, not toy-grade: Single command, sensible defaults, real output bundle, error recovery.
License
MIT — see LICENSE.
Contributing
These are personal skills, but PRs welcome if you want to:
- Add a new skill (open issue first to discuss scope)
- Improve prompt quality on an existing skill
- Add tests / examples
Open an issue with the skill name in the title.
Acknowledgments
- Anthropic Claude Code — the runtime
- OpenAI gpt-image-2 — image generation
- Google Gemini — multimodal analysis
- 千瓜 · 数英 · 我是产品经理 · fxbaogao — XHS merchant research foundation
Routing cheatsheet
When the agent has multiple candidate skills, route by intent:
Content / media
- Image gen — default to
gpt-image(photorealistic / editorial / product / UI / text-in-image / batches > 5). Switch togemini-imagefor multi-reference image input (background swap, character consistency) OR softer / illustrated / Studio Ghibli / watercolor aesthetics. Both can do every category — pick by the differentiators above - PDF generation from markdown →
apple-pdf - Audio → text →
transcribe
Writing
- ZH long-form, personal experience / methodology / 活人感 →
khazix-writer - ZH long-form, industry analysis / company breakdown / cold judgment →
wandian-writer
Research
- "Should I pick A or B" / verify a claim / decide →
deep-research - "Tell me the full story of X" / understand from zero →
narrative-research
Learning / coaching
- 想在 Claude Code 里被主动带着学点东西(英语/某技术/任意技能),要零考试·零背诵·不挫败的引导式学习 →
learning-coach
Workflow / quality
- Build a game / demo / interactive prototype →
mtc - Reviewing a
PLAN.md/ spec →plan-design-review - "Audit the whole codebase" / "clean up everything" →
codebase-sweep - "Audit and fix" / iterative fix on recent changes →
audit-fix-loop - Autonomously close out PR review feedback + CI →
pr-fix-loop - End-of-session knowledge sync (memory + docs + README) →
neat-freak