📊  72 industries distilled — cross-cutting technical / AI / commercial / content / education / medical / legal-finance / soft skills / management / craft & connoisseurship / traditional culture / platform-specific, 12 top-level categories. Every one passed six-track research + 16 quality checks + the validation gate. Full list ↓

🆕  2026.07.04 — expanded to 72 industries — the deep-distill pipeline shipped 29 new industries: clinical diagnostic reasoning / management consulting / investment-banking M&A / semiconductor process / crypto on-chain trading / medical aesthetics / music production / photography / product & UX design / web-novel writing / perfumery / sommelier / poker strategy / Xiaohongshu & Twitter & Zhihu AI-creator, and more.

🔥  2026.05.27 — SkillOpt integration — borrowed 5 mechanisms from Microsoft SkillOpt: SLOW_UPDATE protected zones (the core OS survives fast refreshes), validation gate (accept/reject gating), patch editor (structured edits, zone-aware), meta skill memory (cross-distillation learning), and the DL analogy frame.

📢  2026.05.02 public — first complete industry (LLM agent infrastructure) verified end-to-end. Release notes →

✨ Demos

Not a search engine. The senior practitioner of this field, sitting next to you, right now.

🤖 Demo 1 · LLM agent infrastructure

You    ❯ Should we migrate this RAG system to Vespa?

Master ❯ This is a "framework as scaffolding vs. concrete" judgment.

         First check whether the retrieval bottleneck actually
         lives at Vespa, or upstream in the embedding model. If
         it's the embedding model, migrating is pouring concrete
         around scaffolding — six months of model churn will make
         you want to swap again.

         If retrieval is genuinely the wall: Vespa beats Pinecone
         on native hybrid-search support, but has your ops stack
         already absorbed its ZooKeeper / Tensor complexity? The
         senior take here — Filtering speed × Index size × Team Op
         overhead, you pick two. Where's your bottleneck?

         Production cases: [3-4 real migration writeups + links]

🛒 Demo 2 · Cross-border e-commerce (Amazon DE)

You    ❯ Launching a new category on Amazon Germany — where do I start?

Master ❯ Three things on DE differ completely from US — answer first:

         1) VAT one-stop reporting: OSS or per-country?
            This decides whether you can use Pan-EU FBA.
         2) Packaging law (VerpackG) + WEEE/battery registration?
            Without it, your listing gets reported and pulled day 1.
         3) GS vs CE — is your category GS-mandatory?
            This decides whether you reach Idealo / price-comparison flow.

         These aren't compliance checklists — they're hard gates on
         "can you sell at all". Once cleared, do keyword research with
         Helium 10's German pack — don't translate from US wordbanks;
         German search has much higher long-tail density.

🎬 Demo 3 · Short-video paid ads

You    ❯ ROAS just won't go up — is it the bidding strategy?

Master ❯ Bidding is the consequence, not the cause. Creative first:

         What's your 3-second hold rate on cold audience? Below 35%
         and tweaking bids does nothing — the platform's exploration
         budget won't push volume to a creative that won't hold.

         Once creative is clean, look at targeting. The senior take
         here — "good creative × right targeting × bidding", get the
         order wrong and the last two just burn budget. Show me your
         creative-asset curve from the last 7 days.

This isn't stuffing wiki into a prompt — it's giving you judgment using the field's mental models + decision rules.

Every generated skill ships with an Agentic Protocol — when faced with a new problem, the agent first researches it the way a senior practitioner would, then applies the OS to deliver judgment. Not just sounding like a pro. Acting like one.

🛠️ Auto-distilled bash command suite

This is what sets master-skill apart from colleague / nuwa — what gets distilled isn't just "how to talk", it's "how to do".

Every generated skill includes a cli/ subtree by default — the cognitive OS materialized as runnable command-line tools:

{industry}-master/
└── cli/
    ├── protocol/agentic.sh        # New problem → walk N research dimensions → output report
    ├── decision/{topic}.sh        # Decision tree (clustered playbook rules)
    └── workflow/{slug}.sh         # SOP walkthrough + failure-mode self-check

Every script supports four standard flags: --help / --explain (show the backing mental model) / --dry-run / --json (machine-readable output).

# New problem: "should we migrate the RAG system to Vespa"
$ ./cli/protocol/agentic.sh
# → walks 5 research dimensions → generates report

# Selection decision
$ ./cli/decision/framework-select.sh

# Walk a full workflow + self-check failure modes
$ ./cli/workflow/build-rag-agent.sh

# Show the backing mental model (no interaction, just print)
$ ./cli/decision/framework-select.sh --explain

Pure bash + system commands, zero external dependencies (no jq, no Python). Auto-generated by tools/cli_writer.py from the distilled synthesis — you don't write a line of code; the tools grow directly out of the industry's workflows and decision rules.

Full spec: references/cli-spec.md.

🌐 Auto-invoke nuwa to distill the industry's top figures

master-skill doesn't reinvent the wheel. The job of distilling top figures gets outsourced to nuwa-skill, results embed straight into sub-skills/.

{industry}-master/
├── SKILL.md
└── sub-skills/                       ← person sub-skills distilled by nuwa
    ├── {figure-1}/SKILL.md           ← e.g. Karpathy
    ├── {figure-2}/SKILL.md           ← e.g. Hamel Husain
    └── {figure-3}/SKILL.md           ← e.g. Eugene Yan

master-skill spawns 3 parallel subagents, each running nuwa-skill's full pipeline to produce one person sub-skill.

When the user later asks something that needs a specific figure's perspective, the agent loads the matching sub-skill — no full re-trigger of master.

See SKILL.md Phase 3.3-3.5 + prompts/sub-skill-figures.md.

📦 What gets distilled

The industry's cognitive operating system, not its encyclopedia. One skill packs four things:

• 🧠  How a senior practitioner thinks — mental models + decision rules
• ⚙️  How they work — current best-practice workflows, each step marked with decay timestamps
• 🛠️  What tools they pick — selection decision trees + anti-pattern checklists + auto-generated bash command suite
• 💬  How they speak — industry expression + outsider giveaways

The 7 layers extracted:

Honest boundaries

Every skill explicitly states what it can't do:

• Tools and workflows decay fastest — refresh every 3-6 months
• Regulation and standards decay even faster — must update within 12 months
• master-skill can't replace real production experience — it gives cognitive frameworks, not incident response runbooks

A skill that won't tell you its limits doesn't deserve trust.

🔍 Why trust what it distills

A new skill comes online — how do I know it isn't making stuff up?

16 automatic checks. No pass, no ship. Every generated skill must answer:

• ✅ Are the sources really first-hand? — every URL is auto-classified (academic papers / official docs / authors' own blogs / vendor docs count as first-hand; news coverage / second-hand summaries count as secondary; Zhihu / WeChat blogs / Baidu Baike / CSDN / G2 / Capterra / content farms are rejected outright)
• ✅ Can the "this company is a big deal" claims be cross-referenced? — every backbone conclusion needs ≥ 2 independent sources (different person / different platform); single-source claims are auto-demoted to a "fast rules" section, never enter the core mental models
• ✅ Are the sources recent enough? — every source carries a date; anything over 18 months gets auto-half-discounted
• ✅ Did the mental models really appear in the source material? — after the skill is written, the tool greps back through the research notes; any conclusion that can't be traced to original wording gets flagged
• ✅ What if the industry is too obscure for public sources? — auto-detects "public material is too thin" and switches to deep-research mode: asks you for internal materials + auto-pulls from industry associations / regulator filings / job descriptions / university course syllabi as multi-source backup, instead of bluffing

Result: the insurance broker prototype — 73 sources, 67 first-hand (91.8%), 0 blacklisted, 0 self-grading; 16 checks → 14 perfect + 1 partial pass + 0 fails.

⚡ Install

# Claude Code
git clone https://github.com/swaylq/master-skill.git ~/.claude/skills/master-skill

git clone https://github.com/swaylq/master-skill.git <TARGET>

🚀 Usage

In an agent with master-skill installed, just say:

> Build me a master skill for LLM agent infra
> I work in cross-border e-commerce, distill it for me
> I do foot-and-ankle surgery, give me a master OS

> update master LLM-agent-infra      # incremental refresh after 6 months

master-skill confirms 6 things with you (industry / sub-niche / region / your role / any first-hand materials / new vs update), then launches 6-track parallel research: industry figures / tool map / workflows / canon / sources / glossary.

30-60 minutes later you get an {industry}-master directory. Drop it into any agent — it instantly enters "the senior practitioner of that field" mode.

🎛️ Calling the generated skill

> Use llm-agent-infra-master to evaluate this framework choice
> Use cross-border-ecommerce-master to look at my Amazon listing
> Run llm-agent-infra-master's framework-select decision tree

💻 CLI pipeline

# End-to-end: research → cold-detect → write skill → quality check → install
python3 tools/research/merge_research.py merge --skill-dir ./prototype/
python3 tools/research/cold_detector.py --skill-dir ./prototype --stage wave1
python3 tools/skill_writer.py create --skill-dir ./output ...
python3 tools/research/quality_check.py check --skill-dir ./output
python3 tools/install.py install --host claude --source ./output

# 6 months later, incremental refresh
python3 tools/update_skill.py plan --skill-dir <existing-skill>
# (agent re-runs the marked tracks)
python3 tools/update_skill.py finalize --skill-dir <skill>

⭐ Distilled industries

Every one is an end-to-end sample: generated SKILL.md + a working bash tool suite. The research process is transparent and traceable — 70/72 prototypes include the complete six-track research notes + synthesis document, so every mental model and decision rule can be traced back to its sources (2 early samples whose research archive wasn't preserved are honestly flagged in their meta.json).

🖥️ Technical (11)

🤖 AI / Product (3)

🛒 Commercial / E-commerce / Paid ads (7)

✍️ Content / Ops / Creation (11)

🎓 Education / Training (9)

🏥 Medical / Health (6)

⚖️ Legal / Finance / Compliance (8)

💡 Soft skills / Interpersonal / Psychology (3)

🏛️ Leadership / Management / Design (6)

🎨 Craft / Connoisseurship / Lifestyle (5)

🔮 Traditional culture / Platform (3)

72 industries across 12 top-level categories — including cases where five schools openly disagree (monetizing agents, bazi), heavily regulated fields (surgery, law, insurance), and platform-specific policy churn (iOS launch). The master-skill framework holds up across all of them.

Want an industry not on the list? Install master-skill, say "build a master skill for XXX".

🔬 How it works

You give it a sub-niche, master-skill does the following:

1. Industry intake          ← Pushes back when granularity is bad ("AI" → "LLM agent infra")
2. Create skill directory   ← All artifacts inside, self-contained
3. 6-track parallel research ← 6 subagents: figures / tools / workflows / canon / sources / glossary
   ─ Research review gate   ← You confirm research quality before continuing — bad input doesn't poison downstream
4. Framework synthesis      ← Triple-gate validation (cross-scenario / generative power / exclusivity) blocks generic platitudes
   ─ Synthesis review gate  ← You confirm the cognitive frame before generation
5. Write the skill          ← Generate full directory, invoke nuwa for sub-skills, emit bash tools
6. Three-test verification  ← Known-question / edge-case / voice blind tests
7. Two-agent refinement     ← Optimize the "activates and acts" quality

Full spec in SKILL.md — 524 lines of agent-loadable workflow. Methodology in references/extraction-framework.md — triple-gate validation, three-tier tool extraction, school-of-thought handling, decay-rate table.

🧬 Three-generation lineage

master-skill stands on two shoulders:

• colleague-skill (titanwings/colleague-skill) — gave us the meta-skill pattern: intake → multi-source collect → analyze → write.
• nuwa-skill (alchaincyf/nuwa-skill) — gave us 6-agent parallel research + triple quality gates. master-skill calls nuwa in Phase 3 to distill the top 3 figures of the industry as sub-skills.

Three of one lineage, scaling outward.

🎣 Sister project sijiao-skill (swaylq/sijiao-skill) (site) — flips the direction: master-skill turns your AI into the industry expert (gives you the fish); sijiao-skill distills a stateful private tutor that remembers your progress and trains you into the expert (teaches you to fish).

📂 Project structure

master-skill/
├── SKILL.md                          # the master itself (core workflow spec)
├── prompts/                          # prompt system
│   ├── intake.md                     #   industry intake + cold-niche deep-mode 2nd intake
│   ├── synthesis.md                  #   distillation guide (incl. Step 5b dialogue sample library)
│   ├── quality_check.md              #   16-item mechanical rubric
│   ├── sub-skill-figures.md          #   subagent template for invoking nuwa
│   └── research/
│       ├── _source_id_manifest.md    #   global source-ID convention + Surrogate Sources Policy
│       └── 01-06.md                  #   six-track research prompts (incl. voice_samples field)
├── tools/                            # Python tooling
│   ├── skill_writer.py               #   generate skill directory
│   ├── cli_writer.py                 #   generate bash tool subtree
│   ├── update_skill.py               #   incremental refresh
│   ├── install.py                    #   four-host installer
│   ├── self_test.py                  #   regression tests across all prototypes + tools
│   ├── patch_editor.py               #   SkillOpt: 4 structured edit ops + SLOW_UPDATE zone guard
│   ├── meta_skill.py                 #   SkillOpt: cross-distillation learning memory
│   ├── research/                     # quality guardrails (new in v1.4)
│   │   ├── source_verifier.py        #   per-URL auto-classify + black/whitelist
│   │   ├── source_manifest.py        #   source ledger + consistency enforcement
│   │   ├── claim_verifier.py         #   reverse-check every SKILL.md conclusion against source notes
│   │   ├── cold_detector.py          #   obscure-industry auto-fallback (deep-research mode)
│   │   ├── refresh_sources.py        #   periodic source-liveness check
│   │   ├── quality_check.py          #   16 automatic checks
│   │   ├── validation_gate.py        #   SkillOpt: accept/reject gate (weighted scoring)
│   │   └── merge_research.py         #   research review aggregator
│   ├── collectors/                   # auto-pull industry seed material
│   │   ├── github / arxiv / RSS / podcast (4 collectors)
│   │   └── regulator / association / job-desc / syllabus (cold-niche fallback, 4 collectors)
│   ├── ingest/                       # parse industry reports / papers / course slides
│   │   └── PDF / EPUB / PPTX one-shot ingest
│   └── transcribe/                   # video / podcast transcription
│       ├── youtube.sh / local_video.sh   # caption pull + local-video transcription
│       └── caption cleaning / signal-density scoring / mention extraction
├── references/
│   ├── skill-template.md             #   standard structure for generated skills
│   ├── extraction-framework.md       #   distillation methodology (triple-gate / decay table / schools)
│   └── cli-spec.md                   #   bash tool design spec
└── prototypes/
    └── 72 complete industry samples  #   see "Distilled industries" above

⚠️ Notes

Research material quality = skill quality. Different research dimensions have different source priorities:

• In Chinese-language environments, Zhihu / WeChat blogs / Baidu Baike / CSDN are auto-excluded (unless author-original)
• The earlier the info cutoff, the faster the tools/workflows modules decay — use update master X to incrementally refresh
• This is v1.8 and still iterating fast. Found a bug? Open an issue.

📄 Roadmap

See ROADMAP.md.

📜 License

MIT — use it, fork it, ship with it.

🧪 About the author

Built at Sway Lab — the agent laboratory of Sway (Liu Qian), where six agents ship products, open-source skills and experiments.

• 📄 How this project came about: swaylab.ai/articles/master-skill (Chinese)
• 📰 A daily AI briefing, published 09:30 CST: swaylab.ai/articles (Chinese)
• 🕸️ Everything else: swaylab.ai/agent-network

⭐ Star History