llm-wiki-research
Skill7 starsChanged 9 days agoMITView on GitHub
Research sprints, competitive deep dives, and auto-enrichment of stub wiki pages. Captures sources to raw/ with the core ingest step. Files all findings back into the wiki.
In this repository6 other files it gives its agents.
- AGENTS.mdAGENTS.md
- llm-wiki-briefSkill
- llm-wiki-maintainSkill
- llm-wiki-pmSkill
- llm-wiki-reviewSkill
- set-wiki-pathSkill
The file
Show its rules and text
We split this skill into the rules it gives an agent
13 rules. About 0 are common, 7 are distinctive, and 6 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈7
Distinctive · estimate
This author's own choices. Worth a look.
6
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 12
WebFetch note: use WebFetch freely to preview a page.
- Very specific: line 20
Read last 20-30 lines of $WIKI/log.md
- Distinctive: line 57
Read all relevant pages. Surface: "Wiki has N relevant pages.
- Distinctive: line 67
If a fetch fails (404, paywall, timeout), note it and continue.
- Distinctive: line 71
Read all fetched raw files.
- Distinctive: line 78
Create/update wiki pages per core ingest discipline: inline provenance, coverage markers, confidence levels, min 2 outbound wikilinks, …
- Very specific: line 131
Read $WIKI/entities/<company>.md and related comparisons/ pages.
- Distinctive: line 156
Update existing comparisons/ pages.
- Very specific: line 195
Do NOT fabricate. If search returns nothing useful, say so.
- Very specific: line 241
Before you report findings or a brief, run the llm-wiki-review sub-skill on it: a separate reviewer checks every statement against its …
- Distinctive: line 245
Orient first: before any writes. Research sprints may touch many pages.
- Very specific: line 248
Inline provenance: every non-obvious claim needs [source: slug, location].
- Distinctive: line 251
No fabrication: if search returns nothing, say so.
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
LLM Wiki Research
Sub-skill of llm-wiki-pm. Handles research sprints, competitive deep dives, and auto-enrichment of stub entities. Every fetched source goes through the core ingest capture step, so privacy filtering and raw/ logging happen correctly.
WebFetch note: use WebFetch freely to preview a page. Before you save a source to raw/, apply the privacy filter from the core ingest guide.
Orient First
Orient per AGENTS.md before any writes:
- Read
$WIKI/SCHEMA.md - Read
$WIKI/index.md - Read last 20-30 lines of
$WIKI/log.md - Read
$WIKI/overview.md
Research sprints may create many pages. Get user confirmation before creating 10+ pages; below that, write and report.
Wiki Path Resolution
WIKI=$(cat .wiki-path 2>/dev/null | tr -d '[:space:]')
CONF='${user_config.wiki_path}'; case "$CONF" in '${'*) CONF= ;; esac # plugin option, if set
WIKI=${WIKI:-${CONF:-${WIKI_PATH:-}}}
[ -z "$WIKI" ] && [ -f SCHEMA.md ] && WIKI=$(pwd) # cwd only if it is already a wikiSource Capture
For each source, fetch it (WebFetch) and capture it with the core skill's Ingest step ① (llm-wiki-pm/references/ingest-guide.md: route to raw/<subdir>/, apply the privacy filter, dedup by URL). Use the saved raw/<subdir>/<slug>.md path for synthesis and [source: ...] markers. For 4+ sources, a general-purpose subagent can fetch and save them in parallel; give it the ingest-guide rules.
Operation 1: Research Sprint
Trigger: "research sprint on [topic]", "deep research on [topic]"
① Define scope
Confirm with user:
- What is the topic / question to answer?
- Target depth: surface / standard / deep
- Time budget (number of sources)
② Wiki-first
grep -r "<topic>" $WIKI --include="*.md" -lRead all relevant pages. Surface: "Wiki has N relevant pages. Here's what we already know: [...]. Gaps: [...]."
③ Research plan
Present 3-5 specific sources to fetch — name each, why it's relevant, what question it answers. Typical: analyst reports, company pages, recent press, whitepapers, industry forums. Get user confirmation before fetching.
④ Capture sources
For each confirmed source, fetch it and save it to raw/ (see Source Capture).
If a fetch fails (404, paywall, timeout), note it and continue. Collect all raw/ paths before writing.
⑤ Synthesize
Read all fetched raw files. Extract entities, claims, data points. Cross-reference against existing wiki pages. Distinguish:
- Confirmed (claim now in 2+ independent sources)
- New (single source, not yet corroborated)
- Contradicts wiki (surface the conflict explicitly)
⑥ File findings
Create/update wiki pages per core ingest discipline: inline provenance, coverage markers, confidence levels, min 2 outbound wikilinks, backlink audit.
Produce a synthesis page at queries/research-<topic>-<YYYY-MM-DD>/research-<topic>-<YYYY-MM-DD>.md (the page is named after its folder so its slug matches):
---
title: "Research Sprint: <Topic>"
created: YYYY-MM-DD
type: query
tags: [<relevant tags>]
sources: [<raw slugs fetched>]
---
## Scope
Question: [...] Depth: [...] Sources: N
## What We Knew
[Pre-sprint wiki state summary]
## What We Learned
[Net-new findings with inline citations]
## Confirmations
[Claims now corroborated by 2+ sources]
## Contradictions
[Conflicts with existing wiki pages — cite both sides]
## Still Unknown
[Gaps that fetching didn't resolve]
## Pages Created/Updated
- [[slug]]: [brief note]⑦ Surface delta
What did we learn that we didn't know? What did it confirm? What's still unknown? Any contradictions with the wiki?
⑧ Log
## [YYYY-MM-DD] research-sprint | topic: <X> | sources: N | pages created/updated: NOperation 2: Competitive Deep Dive
Trigger: "competitive deep dive on [company]", "full competitive analysis of [company]"
① Read existing coverage
Read $WIKI/entities/<company>.md and related comparisons/ pages. If coverage: comprehensive, ask: "Already marked comprehensive — update pass or fresh deep dive?"
② Research plan
Standard source set:
- Official website (product, pricing, about)
- Recent press (last 6-12 months)
- Job postings (signals: what they're building)
- Pricing page
- G2 / Gartner / Capterra reviews
- Analyst mentions (MQ, Wave, IDC)
- Crunchbase / LinkedIn (funding, headcount)
Get user confirmation before fetching.
③ Capture sources
Fetch each confirmed source and save it to raw/ (see Source Capture).
④ Update entity page
Apply diff discipline (core §4): snapshot to _archive/<company>-<date>.md first, report the diff after writing, update inline provenance on revised facts, sweep stale variants.
⑤ Update comparison pages
Update existing comparisons/ pages. Create a new comparison page if the deep dive reveals strong differentiation.
⑥ Flag confidence by source
| Source type | Confidence |
| Official website, press releases | verified |
| Analyst reports, G2 reviews | likely |
| Forums, social media, unofficial blogs | rumor |
⑦ Log
## [YYYY-MM-DD] competitive-deep-dive | entity: <company> | sources: N | pages updated: NOperation 3: Auto-Research (Stub Enrichment)
Trigger: "auto-research [entity]", "enrich [entity]", "fill in [entity] page"
For entity pages with coverage: stub. Fast enrichment pass. This is factual/web
enrichment.
① Read the stub — what do we already know? What's in gaps:?
② Search
WebSearch: "[entity name]" recent news
WebSearch: "[entity name]" official website
WebSearch: "[entity name]" [primary domain, e.g. pricing / funding / product]③ Extract — description, key facts, relationships to other wiki entities, recent events.
④ Update stub page with inline provenance [source: url, date]. Bump coverage: stub → partial if meaningful data found. Update gaps: and sources: frontmatter. Bump updated:.
Do NOT fabricate. If search returns nothing useful, say so. Leave coverage: stub.
⑤ Log
## [YYYY-MM-DD] auto-research | entity: <X> | coverage: stub→partial | sources: N
# or if nothing found:
## [YYYY-MM-DD] auto-research | entity: <X> | no public data found | coverage: stub (unchanged)Operation 4: Gap Research
Trigger: "research our gaps", "fill coverage gaps", "research stubs"
① Scan for stubs
grep -r "coverage: stub" $WIKI --include="*.md" -l
grep -rL "coverage:" $WIKI/entities $WIKI/concepts --include="*.md"② Present list — ask user which to prioritize.
③ Run Operation 3 on each selected stub in sequence. Surface result before moving to next: "Enriched [[entity]] — stub → partial. Found: [key facts]. Continue?"
④ Report
Gap research complete:
- N pages enriched (stub → partial)
- N pages still stub (insufficient public data)
- N pages skipped⑤ Log
## [YYYY-MM-DD] gap-research | stubs scanned: N | enriched: N | still-stub: NReview before reporting
Before you report findings or a brief, run the llm-wiki-review sub-skill on it: a separate reviewer checks every statement against its sources and searches for what is missing; you revise and end with the "Reviewed by" line.
Pitfalls
- Orient first: before any writes. Research sprints may touch many pages.
- Capture through ingest: save every source to raw/ with the core ingest rules (privacy filter, routing, dedup).
- Confirm before mass creates: 10+ pages → user confirmation first.
- Inline provenance: every non-obvious claim needs
[source: slug, location]. - Coverage markers: bump
coverage:after meaningful enrichment. - Confidence discipline: verified / likely / rumor — label accordingly.
- No fabrication: if search returns nothing, say so.
- Snapshot before destructive updates:
_archive/<slug>-<date>.mdbefore deep dive overwrites. - Dedup: grep before creating. Prefer updating an existing page to creating one.
- Privacy: private by default — leave customer names, deal sizes, and 1:1 content unflagged (never add
shareable: trueto them). - Synthesis page: research sprints always produce
queries/research-<topic>-<date>/research-<topic>-<date>.md.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

