Knowledge Librarian — turn a dump of files into useful context

Purpose

You have years of existing knowledge sitting in docs, decks, notes, and emails. This skill absorbs it. The user drops material into inbox/; you read, classify, and file it into the Context System so research, strategy, comms, and ops all get smarter. This is the bulk-ingest companion to /brain-dump (which is for quick, in-the-moment verbal capture).

Operating principles applied

• This is the Ingest operation of the Living Wiki (.claude/WIKI.md): read source → synthesize into pages → cross-reference → flag contradictions → archive original to raw/ → update index.md + log.md. You don't just file documents, you build the wiki.
• Principle 2 (brain-dump habitually) & 5 (capture primers): this is how the user's accumulated context stops living only in scattered files.
• Principle 4 (separate planning from execution): you always present a filing plan and get approval BEFORE moving or writing anything.
• See .claude/OPERATING_PRINCIPLES.md and the Wiki Contract in .claude/WIKI.md.

Step 1 — Find the material

1. Determine the source. Default: everything in inbox/ (ignore inbox/README.md). If the user named specific files/folders, use those.
2. ls/Glob the source and report what you found: file count, types, rough sizes.
3. If the inbox is empty, tell the user to drop files into inbox/ (point them at inbox/README.md) and stop. Don't invent material.

Step 2 — Interview the user (only what's needed)

Keep this light — the librarian does the heavy lifting. Use AskUserQuestion:

• Scope — "Process the whole inbox, or just specific files?" (if ambiguous)
• After filing — "Archive processed originals to raw/ (the default), or leave them in place?"
• Sensitivity — "Anything in here I should NOT persist (e.g. confidential legal/HR)?" so the librarian can skip or quarantine it. Pull from existing context/ so the librarian merges instead of duplicating.

Step 3 — Classify & synthesize a plan (spawn the Knowledge Librarian)

Spawn the knowledge-librarian subagent (Agent tool, subagent_type: knowledge-librarian) with the file list and the user's scope/sensitivity answers. It first orients on context/index.md, then returns a FILING PLAN without writing yet. For a large pile, spawn several librarians over disjoint batches in parallel and merge their plans. The plan is a table: source → target wiki page(s) → what gets written/merged → new [[links]] → any contradiction flagged. It synthesizes into existing pages rather than creating duplicates.

Step 4 — Show the plan, get the call (intervention point)

Present the filing plan to the user as a clean table. Call out:

• New pages vs. merges into existing pages, and the [[links]] being added.
• Contradictions the librarian flagged (e.g. a newer org chart vs. the current company.md) — show both versions + dates and let the user decide.
• Items it was unsure about (defaulted to context/reference/). Let the user adjust ("don't file X", "that note is about Jane not John", "merge these two"). Iterate until they approve. Do not write before approval.

Step 5 — Execute the ingest (synthesize → link → provenance → index → log)

On approval, have the librarian execute the plan per the Wiki Contract:

• Synthesize durable signal into the right pages (company.md, profile.md, reference/, one file per person in stakeholders/, one per decision in decisions/). Keep writing-samples/ verbatim and brain-dumps/ raw.
• Never overwrite — merge under dated, sourced headings; add [[links]] to related pages.
• Provenance: archive each processed original to raw/<date>-<slug>.<ext> (immutable) and cite it on the pages it fed.
• Update context/index.md (catalog entries for new/changed pages) and append an INGEST: line per source to context/log.md with the date (date "+%F %H:%M"), pages touched, and raw/ source.

Step 6 — Report

Tell the user exactly what happened: pages created/merged, [[links]] added, any contradictions to resolve, the raw/ archive, and confirmation that index.md + log.md were updated. Suggest next steps — e.g. "Your [[company]] now has the competitive section; /research-analyst and /strategic-advisor will use it automatically." If lots of writing samples landed, suggest /comms-expert to refresh the voice profile. If you flagged contradictions, suggest a /wiki-lint pass.

Pro tips

• Re-runnable: keep dropping new material in inbox/ and re-running — it's a living intake pipe, not a one-time import.
• Big migrations: for hundreds of files, the chief-of-staff can drive this via the shipped multi-model-research-style workflow pattern; mention it.
• Speak the gaps: after sorting, if profile/company still look thin, nudge the user to /onboard or /brain-dump to fill what no document captured.