| name | expand-llm-wiki |
| description | Expand and maintain the ARA LLM Wiki at `research/wiki/`. Use when Codex is asked to backfill, flourish, grow, ingest into, clean up, or improve the LLM wiki knowledge base; create or update entity/concept/theme pages; add OKF-native wiki content; reconcile new daily digests or model tickets into wiki pages; improve wiki cross-links, aliases, sources, index.md, log.md, or generated index.json. |
Expand LLM Wiki
Use this skill to grow ARA's compounding LLM wiki without corrupting its graph.
The wiki is not regenerated. It is CRUD-maintained: update existing pages by
default, create new pages only when the subject is genuinely missing, and never
delete pages.
First Principles Check
Before writing, answer these briefly for yourself:
- What is actually being asked: latest digest ingest, topic backfill, graph
cleanup, schema repair, or OKF export readiness?
- What are the mechanics: Markdown pages under
research/wiki/, strict
frontmatter, alias-aware [[wikilinks]], generated index.json, append-only
log.md.
- What is known directly: facts already in the daily digest, model tickets, or
cited repo sources.
- What is assumed: unsourced claims, vendor benchmarks, rumored timelines, or
anything only inferred from community chatter.
Challenge the premise if the requested expansion would duplicate pages, read
the raw firehose instead of curated synthesis, or turn weak signal into a
permanent wiki page.
Required Context
Read these before editing:
docs/wiki-schema.md in full. It is the contract.
research/wiki/index.md and research/wiki/log.md.
- The relevant existing pages under
research/wiki/{entities,concepts,themes}/.
- The curated input for the task:
- Latest or specified
research/digest/*-digest.md.
- Relevant
research/models/tickets/*.md, especially recently updated
tickets.
- For an explicit topic backfill, any repo-local research files the user
points at.
Do not crawl research/twitter/, research/rss/, research/community/,
research/blogs/, research/arxiv/, or other raw-source directories for a
normal expansion pass. The digest exists to curate those sources. Use web search
only to corroborate a specific unstable claim or when the user explicitly asks
for current external verification.
Expansion Workflow
-
Rank candidate subjects.
Pull out significant entities, concepts, and themes. Ignore passing mentions.
For manual backfills, prefer 4-8 high-leverage subjects over shallow coverage.
-
Search before every write.
Run uv run python scripts/wiki_search.py "<name>" for the canonical name and
likely aliases. If a plausible page exists, update it. Add aliases when a new
name refers to the same subject. Create a page only when no existing page
covers the identity.
-
Update in place.
Fold new facts into the body with concise analysis:
- why it matters
- what changed
- how it links to other pages
- uncertainty and open questions
Bump timestamp to the meaningful update date as
YYYY-MM-DDT00:00:00Z. Do not change created_at, slug, or filename.
-
Create new pages conservatively.
Choose the correct subdir and type: entities, concepts, or themes.
Use stable lowercase kebab slugs. Write OKF-native frontmatter with
description and timestamp, include repo-local sources, and link to
related existing pages. Keep vendor/community claims explicitly caveated.
-
Maintain the graph.
Cross-link generously but intentionally. Add reciprocal links when they make
the receiving page more useful, not merely to silence advisory lint.
Preserve alias uniqueness. Never introduce unresolved [[links]].
-
Update the catalog and log.
Add new pages to research/wiki/index.md under the correct section.
Append exactly one new entry to research/wiki/log.md:
## [YYYY-MM-DD] ingest | Created N (...). Updated M (...).
Keep older log entries unchanged.
-
Regenerate generated artifacts.
Run:
uv run python scripts/build_wiki_index.py
Never hand-edit research/wiki/index.json.
Validation
Run these before reporting completion:
uv run python scripts/check_wiki.py
uv run python scripts/build_wiki_index.py --check
uv run python scripts/export_wiki_okf.py --out /tmp/ara-okf-check --clean
git diff --check
For substantive changes, also run:
uv run python -m unittest discover -s scripts -p 'test_*.py'
cd dashboard && bun run build
Run advisory lint after hard validation:
uv run python scripts/check_wiki.py --lint
Treat lint as guidance. Fix easy high-value reciprocal links, orphans, and stale
pages, but do not churn the graph solely to reach zero warnings.
Quality Bar
- Preserve one page per subject. Duplicate pages are the main failure mode.
- Prefer durable knowledge over news blurbs.
- State uncertainty labels in prose when sources are partial, vendor-adjacent,
preliminary, or rumor-grade.
- Use concrete numbers and dates from sources.
- Keep body content useful to an analyst or LLM agent: mechanisms, implications,
relationships, and open questions beat decorative prose.
- Write only files needed for the wiki expansion unless the user explicitly asks
for schema/tooling changes.