Signal from Expert

Inputs

Require a corpus (files or folders of the user's thinking), expert, and topic. URLs are optional. Ask for missing inputs in one plain-text question.

Workflow

1. Read the corpus

Read every named file in full with cat -n. Do not read other files for context.

2. Choose sources

Check <repo>/essays/<expert-slug>/ for saved pieces. Load the deepapi skill; if the key is unset, use your configured credential-loading method. Make 5+ separate POST /v1/search/web calls, varying <expert> <topic> queries across essays, talks, interviews, and specific sub-questions. Pick the 5–8 most relevant pieces and merge with the user's URLs and saved sources.

Show titles and URLs, one per line, and ask "go?". Wait for approval before scraping.

3. Scrape and save

From the skill directory, fetch all URLs in one batch:

python3 scripts/fetch-sources.py --expert "Paul Graham" --out <repo>/essays/paul-graham \
  https://paulgraham.com/startupideas.html https://paulgraham.com/schlep.html

The helper saves each page as NN-slug.md with a header and verbatim text. Check every head/tail preview for the real first and last lines. Remove leftover layout junk without changing the prose.

If <repo>/essays/AGENTS.md is missing, copy assets/essays-AGENTS.md there.

4. Read all sources

Read every saved source in full with cat -n before writing; use its line numbers for citations.

5. Write the analysis

Use this format with 4 numbered items by default. Repeat the item block as needed, then end with the concluding paragraph:

# DD-MM-YYYY — Signal from <Expert>

Corpus: <files>. Sources: `essays/<expert-slug>/` (N pieces). Agent analysis, not the user's words.

## 1. <Short claim; prefix with "Gap:" for a gap>

<Quote the user's relevant words and explain the connection in 1–2 lines.>

> "<Exact expert quote, 1–4 sentences>"

Full section: `essays/<expert-slug>/NN-slug.md:START-END`

**Co-founder read:** <One paragraph connecting the findings and what to do next.>

• Include at least one Gap: item showing where the expert's work challenges the user's thinking or reveals something missing, backed by a quote.
• Quote both people exactly; never paraphrase the user's reasoning. Keep expert quotes short and point to full passages in saved files.
• One claim per item. Plain English. No hedging. Exclude private personal matters unrelated to the topic.

6. Save and show

Save as <corpus folder>/signal-<expert-slug>.md (a single file's parent folder). Show the full analysis in chat. Do not commit.

Failures

• Truncated source: rerun with a higher --max-chars.
• Poor search results: ask for URLs instead of guessing.
• Fewer than 3 sources: say so and ask; do not pad with weak pieces.