Brain Ingest

Brain Ingest (paste → know)

User does one thing: paste a link, path, script, or blob.
Agent runs the full loop, writes/updates a dossier, and returns what we know + a 3-line so what.

paste → route → extract → CAPTURE (always)
              → distill claims
              → COMMIT (gates; fail → stay inbox)
              → DOSSIER write/update (trusted only)
              → report + so what

Default mode: auto (no extra questions unless blocked).
Cap: 12 commit attempts per paste (rest stay inbox as candidates).

Portability

This bundle ships with references/ (dossier + report templates, extractor notes) — keep that folder next to SKILL.md.

• Hermes: full pipeline. requires_skills above must be installed; uses web_extract and fact_store when available.
• Other agents (Claude Code, Codex, …): standalone mode. No prerequisite skills needed. Use any web-fetch/read tool for extraction, keep the same vault layout on disk (inbox/, knowledge/dossiers/), and replace fact_store commits with dossier bullets only. Everything else — routing, source packs, claim gating, dossier merge, the report shape — works as written.

When to use

• Message is mostly a URL, github.com/..., YouTube link, local path, or “learn this:” + paste
• User says point the brain at X / ingest / absorb / remember this page
• After simple-second-brain exists (bootstrap vault if needed)

Do not use

• Pure chat questions with no source material
• Private credentials, .env, keys (refuse commit; warn)
• Huge monorepo without a path scope (ask once for subpath or fail closed to capture-only)

Modes

If user says “just save” → capture-only. If “don’t commit yet” → draft.

Process (pipeline)

0. Bootstrap

Ensure simple-second-brain vault exists, including dossiers:

ROOT="${HERMES_BRAIN:-$HOME/.hermes/hermes-brain}"
mkdir -p "$ROOT"/{inbox/_done,knowledge/dossiers,knowledge/areas,ideas,goals,daily,templates}

1. Route the paste

Detect one primary kind (first match wins):

Write routing decision into the source pack frontmatter: kind:.

2. Extract (reality surface)

Always produce a source pack file under inbox:

inbox/YYYY-MM-DD-HHMM-ingest-<slug>.md

Must include:

---
status: inbox
captured: <ISO>
source: ingest
kind: youtube|web|github|...
origin: <url-or-path-or-paste>
mode: auto|capture-only|draft
---
# Ingest: <title>

## Origin
- <url or path>

## Extract summary
<what was fetched — tools used>

## Raw / excerpts
<blockquote or fenced excerpts with pointers>

## Claim candidates
1. ...
2. ...

Evidence rule: every later COMMIT must cite this file path and/or the origin URL.
If extract fails: still capture the paste + error; stop (no fake claims).

Done when: source pack path exists on disk.

3. Distill claims

From the extract, list atomic claim candidates (default max 12):

• One fact each, ≤20 words preferred
• Tag confidence: high | medium | low
• high = directly stated in source with quote/span
• medium = clear paraphrase of one passage
• low = inference → never auto-commit

Drop marketing fluff, undated superlatives, secrets.

4. Capture is already done

Source pack = capture. Optionally also append one line to daily/YYYY-MM-DD.md under ## Ingests.

5. Commit (auto mode only)

For each high claim (and medium only if user said “be aggressive”):

1. Run verified-brain-commit-ritual gates: evidence / scope / conflict
2. Evidence must point at origin + source pack path (+ quote)
3. COMMIT → prefer fact_store for one-liners; always fold trusted bullets into the dossier in step 6 (do not scatter only into random knowledge/ingest/ files unless the dossier points at them)
4. INBOX on any fail — append to source pack under ## Held (gates failed)

Never commit low. Never commit secrets.

6. Dossier + what we know + so what (mandatory in auto and draft)

Slug: stable topic key from title/repo/video id (kebab-case, ≤60 chars).
Path: $ROOT/knowledge/dossiers/<slug>.md

6a. Write or update the dossier

• If file missing: create from references/dossier-template.md
• If file exists: merge — append new committed bullets under ## What we know (dedupe near-identical lines); add origin under ## Sources if new; leave prior trusted bullets intact
• Body rules:
  • ## What we know = trusted only (this ingest’s commits + prior trusted on this topic from fact_store/knowledge)
  • Optional ## Untrusted / held = clearly labeled; never mixed into What we know
  • Each trusted bullet ends with a handle: (src: origin) and/or fact_store#id and/or source_pack path
  • Frontmatter: status: committed, updated: <ISO>, origins: […url…], slug:
• capture-only mode: skip dossier update (source pack only)

Done when: test -f dossier path and What we know has only trusted lines.

6b. Chat report

Return to user in this shape:

INGEST complete
origin: …
kind: …
source_pack: <abs path>
dossier: <abs path>

COMMITTED (trusted):
- claim … [fact_id or dossier]

HELD in inbox (untrusted):
- claim … (failed: evidence|scope|conflict|low-confidence)

WHAT WE KNOW NOW:
- <mirror dossier ## What we know — trusted only>

SO WHAT:
1. <action or decision implication>
2. <action or “none”>
3. <action or “none”>

Open questions:
- <gaps the source didn't answer>

Rules:

• Do not list held items under WHAT WE KNOW NOW
• SO WHAT is exactly 3 lines, numbered; derived only from trusted bullets (or “none — monitoring only”)
• Prefer concrete next steps (read X, try Y, watch Z) over vibes

Details: references/what-we-know-template.md, references/dossier-template.md.

Extractor cheatsheet

Details: references/extractors.md.

Bounded automation (non-negotiable)

• Idempotent-ish: same origin same day → update same slug or add -2 suffix; don't spam duplicates without checking rg origin in inbox/knowledge
• Observable: always print source_pack path
• Fail closed: weak claims stay inbox
• Cap work: 12 commits; large codebases need a subpath
• No external publish

Verification checklist

• Source pack on disk with origin + kind
• Every COMMIT has gate report + evidence path
• Dossier path exists under knowledge/dossiers/ (auto/draft)
• Dossier ## What we know contains only trusted items
• Chat report includes dossier: abs path + exactly 3 SO WHAT lines
• User only pasted; agent didn't require a form unless blocked

NEVER

• Commit a low-confidence claim, in any mode
• Commit secrets (.env, keys, credentials) — refuse and warn
• Mix held/untrusted bullets into a dossier's ## What we know
• Invent spoken claims for a YouTube video whose transcript failed to fetch
• Report WHAT WE KNOW in chat without the dossier file existing on disk
• Publish or send anything externally — this skill is local-only

Pitfalls

• Committing model summary without quotes from extract
• Treating whole article as one claim
• YouTube without transcript → capture titles only, don't invent spoken claims
• Ingesting secrets from a pasted .env
• Dumping WHAT WE KNOW only in chat without writing the dossier
• Mixing held/untrusted bullets into dossier ## What we know
• SO WHAT longer than 3 lines or based on untrusted claims