Back to Details

signal-from-expert

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese

Signal from Expert

Signal from Expert

Two ingredients make this work. Both are required:
  1. Volume of the user's honest, unfiltered thinking — it exposes exactly how they see something and their current level of understanding.
  2. A world-class expert's body of work — several essays, articles or transcripts, read in full, reviewed against all of his thinking.
The output is relevance and clarity: the expert's exact words, mapped onto the user's exact situation, including where they are wrong.
这项工作需要两个必备要素：
  1. 用户真实、未经过滤的思考内容体量——能精准展现用户对事物的看法及当前认知水平。
  2. 顶尖专家的作品合集——需完整阅读多篇论文、文章或访谈稿，并与用户的所有思考内容进行比对。
输出内容需具备关联性与清晰性：将专家原文与用户实际情况对应，包括用户存在错误的地方。

Inputs

输入信息

The user's message names three things. If one is missing, ask for it in one plain-text question.
  • Corpus — the files or folders of the user's own thinking to analyze (e.g. 
    journal/2026-08-29.md
    , or 
    daily/
    last 14 days). Read exactly that. Nothing more.
  • Expert + topic — e.g. "Paul Graham on what to build", "Bezos on decision making".
  • URLs (optional) — specific pieces he wants included.
用户消息需包含三项内容。若缺失其中一项，以简洁文本提问索要。
  • Corpus（语料库）——用户用于分析的个人思考文件或文件夹（例如 
    journal/2026-08-29.md
    ，或
    daily/
    文件夹中过去14天的内容）。仅读取指定内容，不得额外扩展。
  • 专家+主题——例如“Paul Graham 谈创业方向”、“Bezos 谈决策制定”。
  • URLs（可选）——用户希望纳入分析的特定作品链接。

Workflow

工作流程

1. Read the corpus in full

1. 完整读取语料库

cat -n
every named file. Do not skim. Do not read other files to "add context" — the corpus is the corpus.
cat -n
every named file. Do not skim. Do not read other files to "add context" — the corpus is the corpus.

2. Build the source list

2. 构建来源列表

bash
ls <repo>/essays/<expert-slug>/ 2>/dev/null   # reuse what is already saved
Then find more with DeepAPI (load the 
deepapi
skill; 
source ~/.deepapi/env
if the key is not set). Run 5+ separate 
POST /v1/search/web
calls with different phrasings of 
<expert> <topic>
(essay, talk transcript, interview, "best essays on", specific sub-questions). Pick the 5–8 pieces most relevant to the topic. Merge with the user's URLs and anything already saved.
Show the list (title + URL, one line each) and ask "go?". Do not scrape before the user answers.
bash
ls <repo>/essays/<expert-slug>/ 2>/dev/null   # reuse what is already saved
随后通过DeepAPI查找更多内容（加载
deepapi
技能；若未设置密钥，执行
source ~/.deepapi/env
）。针对
<expert> <topic>
使用不同表述（如论文、演讲文稿、访谈、“最佳相关论文”、特定子问题），发起5次以上独立的
POST /v1/search/web
请求。筛选出5-8篇与主题最相关的作品，与用户提供的URL及已保存内容合并。
展示该列表（每行包含标题+URL），并询问“是否继续？”。在用户答复前不得进行抓取操作。

3. Scrape and save the sources

3. 抓取并保存来源内容

bash
python3 scripts/fetch-sources.py --expert "Paul Graham" --out <repo>/essays/paul-graham \
  https://paulgraham.com/startupideas.html https://paulgraham.com/schlep.html
One DeepAPI call for all URLs. Writes 
NN-slug.md
per page (header + verbatim text, site layout junk stripped) and prints a head/tail preview per file. Check every preview: the text must start at the real first line and end at the real last line. Fix leftovers by hand; never edit the prose itself.
If 
<repo>/essays/AGENTS.md
does not exist, copy 
assets/essays-AGENTS.md
there and add the 
CLAUDE.md
symlink.
bash
python3 scripts/fetch-sources.py --expert "Paul Graham" --out <repo>/essays/paul-graham \
  https://paulgraham.com/startupideas.html https://paulgraham.com/schlep.html
针对所有URL发起一次DeepAPI请求。为每个页面生成
NN-slug.md
文件（包含标题+原文文本，剔除网站布局冗余内容），并打印每个文件的首尾预览。检查所有预览内容：文本必须从真实起始行开始，到真实结束行结束。手动修复残留问题；不得修改原文内容。
若
<repo>/essays/AGENTS.md
不存在，将
assets/essays-AGENTS.md
复制到该位置，并添加
CLAUDE.md
符号链接。

4. Read every source in full

4. 完整读取所有来源内容

cat -n
each saved file so you can cite 
file:line
ranges. Read all of them before writing anything.
cat -n
each saved file so you can cite 
file:line
ranges. Read all of them before writing anything.

5. Write the analysis

5. 撰写分析报告

Default 4 numbered items. Format, exactly:
undefined
默认包含4个编号条目。格式严格遵循以下规范：
undefined

DD-MM-YYYY — Signal from <Expert>

DD-MM-YYYY — Signal from <Expert>

Corpus: <files>. Sources: 
essays/<expert-slug>/
(N pieces). Agent analysis, not the user's words.
Read all N. Here are the K that hit your exact situation, each with <Expert>'s own words and the file:line for the full passage.
Corpus: <files>. Sources: 
essays/<expert-slug>/
(N pieces). Agent analysis, not the user's words.
Read all N. Here are the K that hit your exact situation, each with <Expert>'s own words and the file:line for the full passage.

1. <Short, plain-English claim>

1. <Short, plain-English claim>

<One or two lines: what the user said in the corpus, and why this passage hits it.>
"<Exact quote, 1–4 sentences>"
Full section: 
essays/<expert-slug>/NN-slug.md:START-END
<One or two lines: what the user said in the corpus, and why this passage hits it.>
"<Exact quote, 1–4 sentences>"
Full section: 
essays/<expert-slug>/NN-slug.md:START-END

2. ...

2. ...

Co-founder read: <One paragraph. What the items say together and what to do next.>

Rules for the items:

- **At least one item must be a gap** — a place where the expert's writing shows the user is wrong, or missing something. Mark it: `## 2. Gap: ...`. Back it with a quote like every other item.
- Quotes are the expert's exact words, short. Never long excerpts — the full text is in the saved file, the `file:line` range points to it.
- Quote the user's own corpus words when calling out a pattern or gap. Never paraphrase their reasoning.
- One claim per item. Plain English. No hedging.
- Skip anything in the corpus that is private personal life rather than the topic at hand.
Co-founder read: <One paragraph. What the items say together and what to do next.>

条目规则：

- **至少包含一处认知缺口条目**——即专家内容显示用户存在错误或认知缺失的地方。标记格式为：`## 2. Gap: ...`。与其他条目一样，需附上专家原文引用。
- 引用内容需为专家原文，且简短。不得使用长篇摘录——完整文本已保存于文件中，`file:line`范围可指向对应内容。
- 指出模式或认知缺口时，需引用用户语料库中的原文。不得转述用户的推理内容。
- 每个条目对应一个观点。使用简洁直白的表述，不得含糊其辞。
- 忽略语料库中与当前主题无关的私人生活内容。

6. Save and show

6. 保存并展示结果

Save to 
<corpus folder>/signal-<expert-slug>.md
(if the corpus is one file, its parent folder). Print the full analysis in chat. Do not commit.
将分析结果保存至
<corpus folder>/signal-<expert-slug>.md
（若语料库为单个文件，则保存至其父文件夹）。在聊天窗口中打印完整分析内容。无需提交版本控制。

Failure modes

故障处理

  • Unknown field "url"
     — the scrape body uses 
    urls
    (array). The script does this right.
  • Python 
    IncompleteRead
    on DeepAPI responses     — the sandbox proxy truncates chunked responses to urllib. The script uses curl for that reason; do the same for any ad-hoc calls.
  • Page shows 
    truncated: true
     — re-run with a higher 
    --max-chars
    .
  • Search returns junk — ask the user for URLs instead of guessing.
  • Fewer than 3 sources found — say so and ask; do not pad with weak pieces.
  • Unknown field "url"
    ——抓取请求体需使用
    urls
    （数组格式）。脚本已正确处理此问题。
  • DeepAPI响应出现Python 
    IncompleteRead
    错误    ——沙箱代理截断了urllib的分块响应。因此脚本使用curl处理，临时调用也需采用相同方式。
  • 页面显示
    truncated: true
    ——使用更高的
    --max-chars
    参数重新运行。
  • 搜索结果无效——向用户索要URL，而非自行猜测。
  • 找到的来源少于3个——告知用户并询问；不得用无关内容凑数。