Signal From Expert

Signal from Expert

Read the user's raw thinking in full, find what a specific expert actually wrote that speaks to it, and report only the passages that land — then name where the expert says the user is wrong.

What to get from the user (ask once, only for what's missing)

1. Corpus — the user's own raw thinking to analyse. A file, a folder, a date range of daily notes, or "this conversation". Example: <vault>/Daily/ last 14 days, or a pasted brain-dump.
2. Expert + topic — who, and on what. Example: "Paul Graham on what to build", "Rich Hickey on simplicity", "Charlie Munger on this pricing decision".
3. URLs (optional) — specific essays/talks/threads the user wants included for certain.

Procedure

1 — Read the corpus completely. No skimming. Also check the Obsidian vault ($OBSIDIAN_VAULT_PATH) for related notes via obsidian_search — the user's own past conclusions are part of the corpus. Write a 3–5 line statement of the user's actual situation, position, and open questions. Show it to the user before going further only if the corpus was ambiguous.

2 — Find 5–8 real sources from the expert. Use WebSearch for essays, talks, interviews, long-form threads. Prefer primary writing over summaries of it. Add any URLs the user gave. WebFetch each one and read it in full. If a source turns out irrelevant, drop it and find another — aim for 5–8 genuinely on-point pieces, not 5–8 links.

3 — Match. Pull only the passages that hit the user's exact situation. For each:

• the verbatim quote (short — the load-bearing sentence or two, not a paragraph)
• the source (title + URL, and the section if long)
• one line: how it applies to the user's situation, pointing at the corpus (daily/2026-09-01.md: "…their line about X").

A passage that's merely on-topic doesn't qualify. It has to change how the user should act.

4 — The gap (required). Name at least one place where the expert's work shows the user is wrong, missing something, or asking the wrong question. This is the most valuable part — do not soften it or skip it because the match section already looks useful.

5 — Read-out. 3–5 sentences on what the passages mean together for the user's next move.

Output shape

**Situation** — 2–3 sentences, in the user's own framing.

**Signal** (3–6 items max)
- "<verbatim quote>" — <Source title>, <url>
  → <how it applies, with a corpus pointer>

**Where you're off**
- <the gap. Direct. Quote the expert. Say what the user should reconsider.>

**Together** — 3–5 sentences. The synthesised move.

Rules

• Quotes must be verbatim and real. If you can't fetch a real source, say so — never paraphrase-then-quote or reconstruct a passage from memory.
• Scrape wide, report narrow. A long list of quotes is a failure. 3–6 that matter.
• No "I searched", "I found", "I read 8 essays" narration. Just the signal.
• If the expert genuinely has nothing precise to say about the situation, say that plainly rather than stretching loose passages to fit.
• Save the read-out to the vault (obsidian_capture or a note under Learning/ or the relevant project) so the challenge compounds.