Lambenthan/empiricalwiki / i18n/en/skills/ask
ask
VerifiedAsk the wiki a question, retrieve and synthesize relevant pages, optionally crystallize the answer back into the wiki
/ask
Ask a question to the wiki knowledge base. The LLM reads context_brief.md for global context, retrieves relevant pages, synthesizes an answer with citations. Good answers can be crystallized back into the wiki — written to outputs/ or as new concept/claim pages — so exploration compounds like ingestion does.
Inputs
question: natural-language question (e.g. "What is the core difference between LoRA and Adapter?")--crystallize(optional): if specified, crystallize the answer back into the wiki (default: answer only, no write)--format(optional): output format, defaultmarkdown, options:table/timeline/bullets
Outputs
- Always: terminal output of synthesized answer (with
[[slug]]citations) - If crystallize:
wiki/outputs/{query-slug}.md— query result page (default crystallize target)- or
wiki/concepts/{slug}.md— if the answer reveals a new cross-paper concept - or
wiki/claims/{slug}.md— if the answer surfaces a new verifiable assertion - updated
wiki/graph/edges.jsonl(relationships produced by crystallize) - updated
wiki/index.mdandwiki/log.md
Wiki Interaction
Reads
wiki/graph/context_brief.md— global compressed context (claims, gaps, failed ideas, papers, edges)wiki/index.md— page catalog for locating relevant pageswiki/graph/open_questions.md— open questions, helps identify whether the question touches known gapswiki/papers/*.md— paper pages relevant to the questionwiki/concepts/*.md— concept pages relevant to the questionwiki/claims/*.md— claim pages relevant to the questionwiki/topics/*.md— topic pages relevant to the questionwiki/people/*.md— if the question involves specific researcherswiki/ideas/*.md— if the question involves research ideas or failed ideaswiki/experiments/*.md— if the question involves experiment resultswiki/Summary/*.md— if the question involves domain-wide landscape
Writes (crystallize mode only)
wiki/outputs/{query-slug}.md— CREATE (query result page)wiki/concepts/{slug}.md— CREATE (newly discovered concept) or EDIT (supplement existing concept)wiki/claims/{slug}.md— CREATE (newly discovered assertion) or EDIT (add evidence)wiki/graph/edges.jsonl— APPEND (relationships produced by crystallize)wiki/graph/context_brief.md— REBUILD (if crystallize created new pages)wiki/graph/open_questions.md— REBUILD (if crystallize created new pages)wiki/index.md— EDIT (if crystallize created new pages)wiki/log.md— APPEND
Graph edges created (crystallize only)
output → paper:derived_from(papers cited in the answer)output → concept:derived_from(concepts cited in the answer)output → claim:derived_from(claims cited in the answer)concept → paper:supports(if a new concept is generalized from papers)claim → paper:supports(if a new claim is extracted from papers)
Workflow
Precondition: confirm working directory is the wiki project root (containing wiki/, raw/, tools/).
Set WIKI_ROOT=wiki/.
Step 1: Load Global Context
- Read
wiki/graph/context_brief.md— get compressed snapshot of wiki's current knowledge (claims, gaps, papers, edges) - Read
wiki/graph/open_questions.md— understand known open questions and knowledge gaps - If both are missing, rebuild first:
python3 tools/research_wiki.py rebuild-context-brief wiki/ python3 tools/research_wiki.py rebuild-open-questions wiki/
Step 2: Retrieve Relevant Pages
- Read
wiki/index.md, match relevant slugs against question keywords - Extract claims and papers semantically related to the question from context_brief.md
- Sort by relevance, select top-K pages (K ≤ 15 to avoid exceeding context window)
- Read full content of selected pages
- If the question involves relationships (e.g. "difference between X and Y"), additionally read edges connecting X and Y from
wiki/graph/edges.jsonl
Step 3: Synthesize Answer
- Synthesize an answer to the user's question based on collected page content
- Answer requirements:
- Cited: every key claim must include a
[[slug]]wikilink pointing to its source page - Structured: organize output according to
--formatparameter (markdown / table / timeline / bullets) - Acknowledge uncertainty: clearly flag "insufficient evidence in wiki" for parts with weak support
- Flag knowledge gaps: if the question touches a known gap in open_questions.md, call it out explicitly
- Cite claim confidence: when referencing claims, note their confidence and status
- Cited: every key claim must include a
- If the question exceeds the wiki's current knowledge, honestly say so and suggest:
- which papers to ingest to fill the gap
- possible search directions (arXiv keywords, Semantic Scholar queries)
Step 4: Assess Crystallize Value
- Judge whether the answer is worth writing back to the wiki (make a recommendation even if
--crystallizewas not specified) - Signals that crystallize is worthwhile:
- The answer synthesizes information from multiple papers, forming a new cross-paper insight
- The answer reveals a concept not yet explicitly recorded in the wiki
- The answer surfaces a new verifiable assertion (claim)
- The answer addresses a known gap in open_questions.md
- Signals that crystallize is not worthwhile:
- The answer merely restates the content of a single page
- The question is a simple factual lookup (e.g. "What year was LoRA published?")
- The answer relies primarily on inference rather than wiki evidence
- Append a crystallize recommendation at the end of the answer:
💡 Crystallize recommendation: [worthwhile / not needed] — [reason]
Step 5: Crystallize Back to Wiki (if user confirms or --crystallize was specified)
Choose the crystallize target based on answer content:
Case A — Write to outputs/ (default):
- Generate slug:
python3 tools/research_wiki.py slug "<query-summary>" - Create
wiki/outputs/{query-slug}.md:
Body is the answer content (preserve wikilinks)--- title: "" slug: "" query: "" # original question source_pages: [] # slugs of all pages cited in the answer date_created: YYYY-MM-DD --- - Add a graph edge for each cited source page:
python3 tools/research_wiki.py add-edge wiki/ --from outputs/<slug> --to papers/<source-slug> --type derived_from --evidence "query answer"
Case B — Create new concept:
- If the answer reveals a new concept: create
wiki/concepts/{slug}.mdusing the CLAUDE.md concept template - maturity: emerging
- key_papers: extracted from answer citations
- Add graph edges (concept → papers)
- Append reverse links to relevant paper pages under
## Related
Case C — Create new claim:
- If the answer surfaces a new assertion: create
wiki/claims/{slug}.mdusing the CLAUDE.md claim template - status: proposed (synthesized from query, not direct experimental evidence)
- confidence: set initial value based on strength of cited evidence
- source_papers: extracted from answer citations
- Add graph edges (claim → papers)
- Append reverse links to relevant paper pages under
## Related
Step 6: Update Navigation and Graph (crystallize only)
- index.md: append new page entries under the appropriate category
- log.md:
If not crystallized:python3 tools/research_wiki.py log wiki/ "ask | <question-summary> | crystallized: <target-path>"python3 tools/research_wiki.py log wiki/ "ask | <question-summary> | answer-only" - Rebuild derived graph files (only if crystallize created new pages):
python3 tools/research_wiki.py rebuild-context-brief wiki/ python3 tools/research_wiki.py rebuild-open-questions wiki/
Step 7: Report to User
Output a summary including:
- Number and list of retrieved pages
- Answer (with citations and formatting)
- Knowledge gap annotations (if any)
- Crystallize recommendation or execution result
- Follow-up suggestions (papers recommended for ingestion, related open questions)
Constraints
- No fabrication: answers must be grounded in actual wiki content; do not invent from LLM pre-training knowledge
- Citations must exist: every
[[slug]]must point to a page that actually exists in the wiki - raw/ is read-only: do not modify files under
raw/ - graph/ only via tools: do not hand-edit files under
graph/ - Crystallize requires confirmation: unless the user explicitly specifies
--crystallize, only recommend but do not write - Context limit: retrieve at most 15 pages to stay within context window
- Cite claim confidence: when referencing claims, always note their confidence value and status
- Flag gaps: if the question touches a known gap in open_questions.md, explicitly call it out
- outputs/ frontmatter must include query and source_pages: ensures traceability
Error Handling
- context_brief.md missing: run
python3 tools/research_wiki.py rebuild-context-brief wiki/to rebuild, then retry - wiki is empty: inform the user to first run
/initor/ingestto build the knowledge base - no matching pages: honestly report that no relevant content exists in the wiki, suggest search and ingest directions
- crystallize slug conflict: append a numeric suffix (e.g.
query-result-2) - index.md missing: run
python3 tools/research_wiki.py init wiki/to initialize, then retry
Dependencies
Tools(via Bash)
python3 tools/research_wiki.py slug "<title>"— slug generationpython3 tools/research_wiki.py add-edge wiki/ --from <id> --to <id> --type <type> --evidence "<text>"— add graph edgepython3 tools/research_wiki.py rebuild-context-brief wiki/— rebuild compressed contextpython3 tools/research_wiki.py rebuild-open-questions wiki/— rebuild knowledge gap mappython3 tools/research_wiki.py log wiki/ "<message>"— append log entrypython3 tools/research_wiki.py init wiki/— initialize wiki (fallback)
Skills(via Skill tool)
/ingest— referenced when suggesting the user supplement knowledge
Shared References
.claude/skills/shared-references/citation-verification.md(created in Phase 3)
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| 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: 198k | Updated | Claude CodeCodex | Verified | Install |
| 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: 194k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-debtDietrichGebert/ponytailHarvest every shortcut: comment into one debt ledger, so deferrals get tracked instead of forgotten. One-shot report. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |
| 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: 194k | Updated | Claude CodeCodex | Verified | Install |