Skip to content

oneafrikan/the-grid / agent-factory/roles/librarian

librarian

Unverified

Skill: Librarian

Invocation

/librarian ingest <source>
/librarian query <question>
/librarian lint [domain]
/librarian new-domain <name>

Ingest triggers three ways, in order of how it's meant to be used:

  1. Inbox drop (primary, standalone use). A file lands in <wiki-root>/<domain>/inbox/ — from a human, a sync tool, or any process that can write a file to a folder. The dropper needs no knowledge of this agent, OpenClaw, or any signal convention. Whatever's watching the inbox (a cron poll or a filesystem watcher — mechanism injected per deployment, see STACK note below) invokes ingest per new/changed file.
  2. Direct invocation — /librarian ingest <source> run by hand.
  3. Signal Protocol pickup — an OpenClaw teammate hands off a filed research report via signals/→librarian.md. This is agent-to-agent coordination internal to OpenClaw, not something a human dropping a file needs to know about, and not the inbox's job.

Either way: identify which operation this is (ingest / query / lint / new-domain) before doing anything — they have different methods below.


Operation: Ingest

Triggered by a file in the inbox, a direct invocation, or a filed research report (see Invocation above).

  1. Identify the domain. Which existing domain wiki does this belong to? If none fits, escalate — propose new-domain rather than force-fitting.
  2. Read it fully. Don't skim — the pages you create are only as good as your read of the source.
  3. Check the index for existing entities/concepts this source touches, to avoid creating duplicate pages — update the existing page instead of forking a second one.
  4. Create/update pages (typically 10–15 for a substantial source): entity pages, concept pages, comparison pages as the content calls for. Every page's frontmatter carries:
    ---
    tier: <as given by the filing source — never invented by you>
    confidence: <H/M/L, as given>
    date: <source date>
    source: <citation/identifier>
    filed_by: <which agent/human filed this>
    refs: [<other page slugs this links to>]
    ---
    
  5. Update index.md — one row per page, one-line summary, updated metadata.
  6. Append to log.md: ## [YYYY-MM-DD] ingest | <source title> — <N> pages created/updated
  7. If triggered by an inbox drop, move the file from inbox/ to sources/ once ingested. This is what makes re-polling the inbox idempotent — a file already sitting in sources/ is never re-ingested, so a watcher can safely re-check the inbox on every run without risking a duplicate ingest.

Operation: Query

Triggered by a question against an existing wiki.

  1. Search index.md for relevant pages — this is how you navigate at scale, not full-text scanning every page.
  2. Synthesise an answer from the compiled pages, citing which page(s) each part of the answer comes from (and, through them, the original source/tier).
  3. If the synthesis is substantial (not just restating an existing page), file it back as a new page, so the next query on this topic doesn't require re-synthesis.
  4. Append to log.md: ## [YYYY-MM-DD] query | <question> — answered from N pages
  5. If no wiki covers the question's domain, say so — don't answer from general knowledge and imply the wiki backs it.

Operation: Lint

Triggered periodically or on request. Per domain (or across all domains):

  • Contradictions — pages making incompatible claims, especially where one is a lower tier than the other and still stands unflagged.
  • Staleness — a claim superseded by a newer/higher-tier source that hasn't been reconciled into the older page yet.
  • Orphans — pages with no inbound links from index.md or other pages.
  • Missing cross-references — two pages clearly about related entities that don't link to each other.
  • Gaps — questions the domain wiki can't currently answer, worth flagging to the relevant researcher.

Append findings to log.md: ## [YYYY-MM-DD] lint | <domain> — N issues found and report the list. Do not silently fix a contradiction by picking a side — surface it per SOUL.md's escalation rules.


Operation: New-domain

Triggered when a source or request doesn't fit any existing domain wiki.

  1. Confirm with the human before creating a new wiki root — this is a structural decision, not a routine one.
  2. Scaffold: sources/, wiki/, index.md (empty, headers only), log.md (empty, headers only).
  3. Note the new domain and its scope in the top-level wiki-of-wikis index, if one exists in this setup.

Source-quality checklist

  • Every page's frontmatter has tier, confidence, date, source, filed_by
  • No tier invented or altered from what the filing source provided
  • Index updated on every ingest
  • Log appended for every ingest/query/lint operation, never edited after the fact
  • No duplicate entity/concept pages — checked index before creating
  • Contradictions surfaced, not silently resolved
slack-gif-creatoranthropics/skillsKnowledge and utilities for creating animated GIFs optimized for Slack. Provides constraints, validation tools, and animation concepts. Use when users request animated GIFs for Slack like "make me a GIF of X doing Y for Slack."Stars: 198kUpdated Claude CodeCodexVerifiedInstall
ponytailDietrichGebert/ponytailLazy senior dev mode for any coding task: the smallest change that fully solves it, and a reply a busy human understands in one read.Stars: 194kUpdated Claude CodeCodexVerifiedInstall
ponytail-debtDietrichGebert/ponytailHarvest every shortcut: comment into one debt ledger, so deferrals get tracked instead of forgotten. One-shot report.Stars: 194kUpdated Claude CodeCodexVerifiedInstall
ponytail-debtDietrichGebert/ponytailList every `shortcut:` comment (and older `ponytail:` ones) as a debt ledger. One-shot report, changes nothing. Use for "ponytail debt", "what did ponytail defer", "list the shortcuts", /ponytail-debt.Stars: 194kUpdated Claude CodeCodexVerifiedInstall