Why use it: you already know who you want to distill (Munger, Feynman, Paul Graham, Trương Nhất Minh…) and want to produce a runnable skill with traceable sources.
What it produces: a self-contained skill folder .claude/skills/<tên>-perspective/ comprising SKILL.md, references/research/01-06.md, references/sources/.
⚠️ This entire file is a CHAT command, typed into the agent's conversation window, not the terminal. Shell commands are in 05-helper-scripts-cli.md.
Startup command
蒸馏一个保罗·格雷厄姆
造一个张小龙的视角Skill
distill Paul Graham
What the agent will ask (Phase 0A) and how to answer so you don't get stuck
SKILL.md specifies 6 confirmation items. The most important item is Level (mức chắt lọc) — because this is a long task with many agents and many search rounds; the repo's documentation states outright that a single run on a high-end model can cost tens of USD:
| Level | Research scope | Use when | Cost |
|---|---|---|---|
| Fast (nhanh) | 3 dimensions (著作 + 对话 + 表达), max 5 sources per dimension | testing effectiveness / a person with little material / a tight budget | ≈ 1/3 of the standard level |
| Standard (chuẩn, default) | full 6 dimensions | most cases | medium |
| Deep (sâu) | 6 dimensions + loading all original material (books/subtopics/quotations) | planning to open-source a polished version | highest |
Answer briefly to run the standard level directly:
Nhân vật: Charlie Munger. Chân dung toàn diện. Dùng làm cố vấn tư duy.
Skill mới (chưa có sẵn). Không có tư liệu local. Mức 标准.
Want it cheap:
蒸馏芒格，走快速档，只要3个维度
If you have original material on hand, say so right away — PDF books, interview transcripts, video subtitles, blog exports. The repo calls this Local corpus mode: the agent reads the material you provide first, then only searches the web for the missing dimensions. The quality is markedly higher than pure web search.
Tôi có 2 quyển PDF và 3 file transcript ở ~/tulieu/munger/, ưu tiên dùng chúng.
6-phase pipeline, and 4 stop gates you must approve
| Phase | What the agent does | Deliverable |
|---|---|---|
| 0.5 | Create the skill directory tree before research | references/research/, references/sources/{books,transcripts,articles}/ |
| 1 | Spin up 6 subagents in parallel, each agent one dimension | 01-writings.md … 06-timeline.md |
| 🔴 1.5 | GATE — research quality summary table | table of number of sources / key findings / points of contradiction |
| 2 | Distill the thinking framework (3–7 mental models, 5–10 heuristics, expressive DNA, tensions, honesty boundaries) | distillation notes |
| 🔴 2.5 | GATE — confirm the distillation before writing | summary of N models / N heuristics |
| 3 | Assemble into SKILL.md according to references/skill-template.md |
SKILL.md complete |
| 🔴 4 | LOCK — 3 tests by independent subagents | pass/fail report on 6 criteria |
| 🔴 5 | LOCK — 2 agents refining in parallel | improvement patch |
Six research dimensions in Phase 1:
| # | Dimension | What to hunt for | Output file |
|---|---|---|---|
| 1 | Works | arguments repeated ≥3 times (= true beliefs), self-coined terminology, list of suggested books | 01-writings.md |
| 2 | Dialogue | how they respond when pressed, improvised analogies, moments of changing stance, sentences declining to answer | 02-conversations.md |
| 3 | Expression | high-frequency words/syntax, controversial stances, style of humor | 03-expression-dna.md |
| 4 | The Other | patterns outsiders observe, criticisms, comparisons with peers | 04-external-views.md |
| 5 | Decision-making | background + logic of major decisions, recollections, places where words and actions contradict | 05-decisions.md |
| 6 | Timeline | milestones, turning points in thought, dynamics of the last 12 months (anti-obsolescence) | 06-timeline.md |
Don't skip checkpoint 1.5. The document states clearly: research quality determines the ceiling of skill quality — blocking here is far cheaper than redoing it in Phase 4. If you want to add something, say it directly:
Chiều 05 决策 mỏng quá, chỉ 2 case. Chạy lại agent 5, tìm thêm các quyết định giai đoạn 1990-2000.
Six criteria of Phase 4 (know them in advance to avoid surprises)
| Criterion | Pass | Fail |
|---|---|---|
| Number of mental models | 3–7, each with sourced evidence | <3 or >10 |
| Limitations of each model | clearly state the conditions under which it fails | only list advantages |
| Recognizability of expressive DNA | read 100 characters and recognize who it is | generic ChatGPT voice |
| Boundaries of honesty | ≥3 specific limitations | only "cannot replace the person themselves" |
| Internal tension | ≥2 pairs of contradictions | overly consistent viewpoints (= fake) |
| Ratio of primary sources | >50% | mostly secondary retellings |
These six criteria have a self-check script — see 05-helper-scripts-cli.md. Iteration ceiling: Phase 2→4 loop at most 2 times; if after 2 rounds there are still failing items, record them in the boundaries of honesty and deliver the best available version, without endless polishing.
Generated directory structure
.claude/skills/<person>-perspective/
├── SKILL.md
├── scripts/
└── references/
├── research/
│ ├── 01-writings.md
│ ├── 02-conversations.md
│ ├── 03-expression-dna.md
│ ├── 04-external-views.md
│ ├── 05-decisions.md
│ └── 06-timeline.md
└── sources/{books,transcripts,articles}/
Self-containment rule: every research file must be located inside the skill directory. Copying the whole directory works on another machine — this is the condition for open release.
Verified pitfalls
- Chinese figures: Phase 0.5 automatically switches the source strategy to original Bilibili videos / 小宇宙 podcasts / reputable Chinese newspapers. The source blacklist always includes Zhihu, WeChat Official Accounts, Baidu Baike — all three are excluded in every mode.
- The generated skill's frontmatter description must be < ~300 words, and the skill-loader has a hard cap of 1024 characters. Stuffing in long-tail keywords will trigger an over-limit error, burn tokens every session, and increase false activations.
- Example in
examples/is inconsistent in structure: Jobs/Karpathy/Trump/MrBeast usereferences/research/01-06.md; while Munger/Feynman/Taleb/Musk/Naval use flat, freely named files placed inreferences/.merge_research.pyonly works with the 01-06 structure (verified: running onmunger-perspectivereports❌ 目录不存在and exits 1).