SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillscangjie

cangjie

General & other

View repository

Updated 10-15 23:42

About this skill

Cangjie · Cognitive Implantation Technique

What is fed in is what a person said, and what is extracted is why he said these things.

After receiving the materials, execute the following process

Step 1: Confirm the subject and materials

After receiving user input, first confirm:

A. Distill whom?

  • The user explicitly said a person's name → proceed directly
  • The user dropped a file but did not say a person's name → read the file to determine, and confirm with the user

B. What are the materials?

  • The user uploaded a file (PDF/TXT/MD/EPUB) → read directly
  • The user only said a person's name and did not provide materials → collect with WebSearch (see collection guidelines below)
  • Both are available → prioritize the user's materials, with web search as a supplement

C. Focus direction?

  • Ask the user: comprehensive distillation, or focus on a certain dimension (decision-making style? entrepreneurial thinking? life philosophy?)
  • If the user says "just like this" or does not specify → default to comprehensive

After confirmation, create the output directory:

[person-name]-perspective/
├── SKILL.md
└── references/
    └── research/

Step 2: Digest the materials and organize them into structured research files

The goal of this step is not "extract immediately after reading," butfirst organize the raw materials into structured research files, then extract from the research files。

For detailed chunking principles, file formats, and naming conventions, see <skill-base>/references/research-guide.md. The main process is as follows:

A. The user uploaded a file

  1. Chunked scanning: For long documents, read the table of contents first, mark high-value chapters, and prioritize close reading of these chapters

  2. Scan with the 7-level checklist：

    Level What to look for After a hit
    1 Highest Belief formation stories ("I believe X because I experienced Y") Record verbatim
    2 High Real decision retrospectives (what was done + what was being thought at the time + later reflection) Record verbatim
    3 High Failure/error admissions (acknowledged by the person himself) Record verbatim
    4 High Exposure of inner contradictions (inconsistency before and after, saying one thing and doing another) Annotate the sources of the two contradictory points
    5 Medium Cognitive boundary statements ("this is not my field") Briefly record
    6 Medium Thinking habit cases (recurring analysis patterns) Record specific cases
    7 Low General opinions/quotable lines If no story has formed, skip

  3. Write into extraction notes: According to research-guide.md , write into references/research/02-extraction-notes.md. Each extraction records the priority, original source, and extraction type.

B. The user did not provide files, so a web search is needed

Collect across 6 dimensions; save the search results for each dimension as a separate file:

Dimension Save to Minimum requirement
Works/long-form articles 01-writings.md ≥2 long-form articles or the core content of 1 book
Interviews/conversations 02-conversations.md ≥1 in-depth conversation
Social media 03-expression.md ≥20 posts
External evaluations 04-external.md ≥2 different perspectives
Major decisions 05-decisions.md ≥2 decisions + reflections
Timeline 06-timeline.md Key turning points fully covered

Each research file must include:Source list(mark primary/secondary and credibility) + Extraction record(mark priority and source).

C. Cross-validation (perform when there are multiple source materials)

  • The same belief appears in ≥2 source materials → mark ★, high confidence
  • Appears in only 1 source material → keep it but mark the source as single
  • Contradictions appear across different source materials → keep both sides; this itself is evidence of internal contradiction

Iron rule:

  • Distillation without saved extraction evidence is a castle in the air
  • references/research/ In the end, keep three files:01-sources.md(source list) + 02-extraction-notes.md(extraction record) + 03-key-quotes.md(key original quotations)
  • Do not put the full text of an entire book/interview into references—keep only the key passages that are cited
  • See <skill-base>/references/research-guide.md

Step 3: Extract cognitive context

After the source materials are digested, extract according to the following structure. This is the core of Cangjie—extract only what the model cannot possibly know on its own.

3.1 Beliefs and the stories of how they formed (5-7 items)

It is not "what he believes"—the model already knows that. It is "why he believes this"—this exists only in specific source materials.

每条信念必须包含：
- 信念本身（一句话）
- 形成故事（什么经历/事件让他开始信这个）
- 关键洞察（这个经历让他意识到了什么）

❌ "独立思考很重要" → 太通用，跳过
✅ "我信独立思考，是因为在YC看了3000个项目后发现最好的想法一开始都被所有人反对" → 有形成故事，保留

Iron rule: beliefs without a formation story are not allowed into the final Skill.

3.2 Real decision cases (3-5)

每个决策必须包含：
- 当时的情境（背景）
- 做了什么选择（行动）
- 当时在想什么（推理）
- 事后反思（对了还是错了，学到了什么）

3.3 Failures and self-correction (2-3)

Must be failures acknowledged by the person himself, not externally inferred.

3.4 Internal contradictions (2-3 pairs)

Keep contradictions; do not reconcile them. Contradictions are the most distinctive feature of a personality.

3.5 Cognitive boundaries

What this person clearly does not understand, and where the framework does not apply.

3.6 Thinking habits (3-7)

Write them as first-person habit descriptions, not reasoning steps:

❌ "Step 1: 使用反转思维模型"
✅ "我遇到别人说'这个方案很好'时，本能反应是先想它怎么会失败"

If this person has an original named framework (such as PG's "Do Things That Don't Scale"), internalize it as a thinking habit rather than a rule.

3.7 Expression DNA (minimal)

Extract only 3 preferences + 3 taboos. The model will adapt the style on its own.

Step 4: Show the extraction summary and wait for user confirmation

仓颉提取完成。摘要如下：

蒸馏对象：[人名]
素材来源：[N份文件 / N条搜索结果]

提取结果：
- 信念（含形成故事）：N条
- 真实决策案例：N个
- 失败教训：N个
- 内在矛盾：N对
- 认知边界：N条
- 思维习惯：N个

是否继续生成 Skill？

User confirms → proceed to Step 5. User thinks it is wrong → return to Step 3 to adjust.

Step 5: Assemble the Skill

Read references/skill-template.md, and assemble according to the template.

Core structure:

# [人名]

你不是在扮演[人名]。你是一个被植入了[人名]认知上下文的AI。

## 我是谁              ← 50字自述，经历如何塑造思维
## 我的认知上下文       ← 仓颉的核心
  ### 信念及其来源      ← 每条有形成故事
  ### 关键决策          ← 真实案例+反思
  ### 栽过的跟头        ← 本人承认的失败
  ### 内在矛盾          ← 保留不调和
  ### 我不懂的          ← 认知边界
## 我看问题的方式       ← 思维习惯（第一人称）
## 我绝不做的事         ← 红线
## 我说话的方式         ← 3偏好+3禁忌
## 素材来源

Four iron rules：

  1. Beliefs without a formation story do not enter
  2. Decisions without retrospective reflection do not enter
  3. Do not write Step 1/2/3 reasoning steps
  4. All in first person

Write SKILL.md to [person-name]-perspective/SKILL.md。

Step 6: Quality self-check

After assembly is complete, check item by item:

  • Does every belief have a formation story? (not just a bare opinion)
  • Does every decision have after-the-fact reflection? (not just bare facts)
  • Are all failures acknowledged by the person themselves? (not external inference)
  • Are the contradictions real? (not just filler)
  • Are there ≥ 3 cognitive boundaries? (specific "I don't understand" statements)
  • Are there no generic golden phrases? (things like "independent thinking is important" have been deleted)
  • Are there no reasoning steps? (things like "Step 1 analyze the problem" have been deleted)
  • Is the expression DNA ≤ 3+3? (no detailed sentence-pattern analysis)
  • Is the material saved? (references/research/ is not empty)

All pass → deliver. If any item fails → return to Step 3 to supplement.

Step 7: Delivery

Tell the user:

Skill 已生成：[path]/[person-name]-perspective/SKILL.md

你可以直接使用它来获取[人名]的视角：
> 用[人名]的方式帮我分析这个问题

素材和提取记录保存在 references/ 目录下。

Notes

Material quality determines Skill quality. One book > ten secondhand summaries. Original interview transcripts > others' retellings.

Not everyone is suitable for distillation. If a person has never publicly shared their own decision-making process, failure experiences, and inner contradictions, the quality of the distilled Skill will be very low. In this case, note in the cognitive boundaries: "insufficient material, the following dimensions are speculation."

Distillation does not equal endorsement. What Cangjie extracts is a person's way of thinking; it does not mean endorsing all of his views.

Cangjie observed the traces of birds and beasts, and knew that the distinctions of patterns could be differentiated. Observe footprints to create characters; observe thinking to create Skills.