Wiki research (orchestrator)

Naming: This skill is wiki-research (ad-hoc topic). Use wiki-research-loop only for recurring batch runs from research_loop.tasks_file.

When to use

• User gives a research question, topic, or URL (not a predeclared JSON task list).
• You need to discover the right fetch strategy, pull content into raw/, then merge into wiki/.

Pre-flight check

Before classifying the request, verify the vault is ready. If llm-wiki/.agent-memory.md exists, skim Pitfalls and Preferences (wiki-learn) before heavy research. If memory.enabled, optionally llm-wiki memory recall "<topic>" to see whether past sessions covered this topic (wiki-session-memory).

python3 -c "
import json, pathlib, sys
p = pathlib.Path('llm-wiki/config.json')
if not p.exists():
    print('PRE_FLIGHT: FAIL — config.json not found')
    sys.exit(1)
meta = json.loads(p.read_text()).get('_meta', {})
if not meta.get('setup_completed'):
    print('PRE_FLIGHT: WARN — setup not completed')
else:
    print('PRE_FLIGHT: OK')
"

On FAIL: Stop. Offer to run wiki-setup now:

The llm-wiki vault is not initialized. Would you like me to run the setup wizard?

• [1] Yes — run wiki-setup
• [2] No — cancel

On WARN: Inform the user and ask:

The vault exists but setup hasn't been completed. Some integrations may not work.

• [1] Continue anyway
• [2] Run wiki-setup first

On OK: Proceed directly to Step 1.

See skills/references/preflight.md for the full pre-flight pattern and integration-specific checks.

Step 1 — Classify the request

Read $ARGUMENTS (or the user's message) and assign a primary modality:

When ambiguous, default to wiki-research-web for URLs and wiki-research-deep for open-ended topics.

Open skills/wiki-research/references/approaches.md to see the full dispatch table with example triggers.

Step 2 — Dispatch to sub-skill

Invoke the matched sub-skill and follow its instructions exactly. Pass the original $ARGUMENTS and any relevant context (vault paths, config) forward.

Sub-skill reference files:

• skills/wiki-research/references/approaches.md — full dispatch table and decision tree
• skills/wiki-research/references/query-design.md — how to construct queries per modality
• skills/wiki-research/references/source-eval.md — scoring, dedup, prompt-injection checks
• skills/wiki-research/references/synthesis.md — merging multi-source findings into wiki pages

Step 3 — Post-process (always runs)

Regardless of which sub-skill ran:

1. Validate raw files — llm-wiki raw validate on any new files in raw/
2. Merge into wiki — invoke wiki-ingest and wiki-maintainer
3. Log — append wiki/log.md:

   ## [YYYY-MM-DD] research | <short topic>
   - Modality: <web|academic|social|feeds|news|deep>
   - Sources: <paths under raw/>
   - Pages touched: <wiki/ paths>
   - Open questions: …

4. Optional git snapshot — llm-wiki git snapshot -m "research: <topic> [ingest]" --phase ingest then --phase wiki after merge (if git.enabled)

Done looks like

• Pre-flight OK (or user explicitly continued after WARN).
• Sub-skill completed; llm-wiki raw validate run on new raw/ files; wiki-ingest and wiki-maintainer applied; wiki/log.md has a dated research | … entry with modality, sources, pages touched, and open questions.
• Optional: vault git snapshots with phase tags when git.enabled.

Artifacts

Related skills

• wiki-research-web — URL / web page fetching
• wiki-research-academic — arXiv, DOI, Semantic Scholar
• wiki-research-social — Twitter/X, Hacker News, Reddit
• wiki-research-feeds — RSS / Atom / newsletters
• wiki-research-news — current events via Perplexity or news APIs
• wiki-research-deep — multi-pass, multi-source synthesis
• wiki-research-loop — batch tasks from research-tasks.json
• wiki-extract-youtube — YouTube transcripts and captions
• wiki-extract-paywall — bypass paywalled articles (Freedium, archive.ph, 12ft.io, etc.)
• wiki-extract-ebook — extract text from EPUB, PDF, MOBI, AZW3
• wiki-extract-annas — search and download from Anna's Archive
• wiki-ingest — merge raw/ into wiki/
• wiki-maintainer — index, cross-links, consistency
• wiki-raw-prepare — clean structurally broken markdown in raw/
• wiki-session-memory — recall past sessions on the topic (raw/memory/)
• wiki-fetch — single-URL fetch without wiki merge (lightweight)

Optional: skills/references/context-persona.md and persona.name in llm-wiki/config.json (default Gennie).

Smoke check

• CLI: Run llm-wiki integrations status and any llm-wiki line from Step 1 of this skill (from the vault root).
• Prompt: Invoke this skill by name in Claude Code; complete Step 1 only and confirm expected CLI or file output.