| name | digest |
| description | Deep-propagate one or more ingested sources across the wiki — update concept/entity/question pages, flag contradictions, create new pages where warranted. Optional step after ingest to ensure claims ripple through the full wiki. |
Digest source
Goal: take one or more raw files that already have a wiki summary page (created by ingest) and produce the full ripple effect — surgical propagation into existing pages + new pages where genuinely warranted + index/log updates. This is the skill that turns an ingested source into integrated knowledge across the wiki.
When to use
- After running the
ingest skill, as an optional deeper pass.
- When the user says "digest" / "propagate" a source across the wiki.
- When you notice source-summary pages whose claims haven't been woven into concept or entity pages yet.
Preconditions
- At least one source-summary page in
wiki/ whose claims haven't been fully propagated.
- You have read the wiki schema (
CLAUDE.md or AGENTS.md) and wiki/home.md at least once in this session.
Architecture
If your agent supports subagents (e.g., Claude Code's Agent tool), delegate the heavy lifting:
- Identify which sources need digesting.
- Read the wiki schema,
wiki/home.md, and wiki/index.md to understand the current wiki state.
- Dispatch a subagent with a tight brief (see template below).
- Receive the subagent's report and relay it concisely to the user.
Why delegate: the main thread should not hold multiple source bodies simultaneously. The subagent reads sources once, makes surgical edits, and returns a diff summary.
If your agent does not support subagents (e.g., Codex, Cursor), do the work inline — read the schema, identify sources, and make the edits yourself following the same rules in the subagent brief below.
Step 1 — Identify sources to digest
If the user didn't name specific files, check which source-summary pages exist but whose claims don't appear in concept/entity pages:
grep -l "^type:" wiki/*.md
Confirm with the user before digesting if there's ambiguity.
Step 2 — Load the wiki's current state
Read these three files (and pass their content to the subagent, if using one):
CLAUDE.md (or AGENTS.md) — schema, conventions.
wiki/home.md — current through-line and live tensions.
wiki/index.md — full catalog of existing pages.
Step 3 — Dispatch the digest subagent (or do it inline)
If using subagents, launch one with this brief (adapt per run). Otherwise, follow these same instructions yourself:
Job: Digest source(s) into the wiki. Follow the wiki schema (CLAUDE.md / AGENTS.md) exactly.
Sources to digest: <list of wiki source-summary pages>
Current wiki state: wiki/home.md is the human narrative; wiki/index.md is the agent catalog. Read home.md first to ground yourself, then index.md to know what pages exist.
For each source, do the full ripple:
-
Read the source summary and its raw. Once, carefully.
-
Propagate and cross-link. Make surgical edits (not rewrites) to existing pages the source materially informs. Rules:
- Add to existing paragraphs/sections where the new material belongs; don't create duplicate sections.
- Cite the source:
([[<slug>]]).
- If the source contradicts an existing claim, add a
> [!contradiction] callout inline.
- If a claim has a canonical home elsewhere, link rather than duplicate.
- Add backlinks: when you update an existing page to reference a new concept, also edit the new concept page to link back. Every connection should be bidirectional.
- No orphans: after propagation, every source-summary and every concept page must have at least 2 inbound
[[wikilinks]] from other pages. If a page has zero inbound links, find related pages and add references to them.
-
Create new pages only when the material genuinely warrants one — a new entity, concept, open question, or strategic position — not for every tangent. Lean conservative.
-
Update wiki/index.md to list every new page under the right category.
-
Prepend one log entry per source digested to wiki/log.md (below the header): ## [YYYY-MM-DD HH:MM] digest | <title>. Newest entries at top.
Rules:
- Never modify
raw/. Read-only.
- Stay flat. No new subdirectories inside
wiki/ — wiki/sources/ is the only allowed subdirectory.
- Don't invent facts. If the source doesn't say something, don't claim it.
- Don't touch schema files (
CLAUDE.md, AGENTS.md) or home.md unless the source forces a schema change (rare — flag it, don't just do it).
- Don't rewrite existing pages wholesale. Surgical edits only.
- Don't duplicate content across pages. One canonical home per claim, others link.
Deliverable (under 300 words):
- New pages created (filenames only).
- Existing pages updated (filenames only).
- Any contradictions flagged.
- Judgment calls worth the user reviewing.
- Any new open questions the wiki should track.
Step 4 — Relay the subagent's report
Summarize the subagent's return concisely (under 200 words). Highlight:
- Anything flagged as a judgment call.
- Any contradictions.
- Any new open questions surfaced.
Then ask the user if anything needs adjustment.
Batching multiple sources
If multiple sources need digesting and you're using subagents, dispatch one subagent for all of them rather than one per source. Cross-references between sources stay coherent that way.
Rules
- Never skip Step 2. You (or the subagent) need that context.
- Never run this on a source without an existing summary page — run
ingest first.
- If using subagents: never pre-read raws in the main thread before dispatching. The subagent reads them. Prefer one subagent for all sources over N subagents for N sources.