From llm-wiki-agent
Use when planning or brainstorming a non-trivial task and you want a thorough, one-shot sweep of the agent's past learnings before committing to an approach — best run as a dispatched subagent so the scan happens in its own token budget and returns only a synthesized, cited brief. Read-only. For quick interactive lookups during a conversation, use learn-recall instead.
How this skill is triggered — by the user, by Claude, or both
Slash command
/llm-wiki-agent:learn-researchThis skill is limited to the following tools:
The summary Claude sees in its skill listing — used to decide when to auto-load this skill
Sweep the compound-learnings store at **planning time** and return a short,
Sweep the compound-learnings store at planning time and return a short,
cited brief of the lessons that bear on the task at hand — past corrections to
respect, playbooks to reuse, insights/patterns to factor in. This is the
heavier, comprehensive counterpart to learn-recall.
Run this as a dispatched subagent. The point is isolation: the subagent does the grepping and body-reading in its own context budget and hands back only the synthesized findings, so the planner's context stays lean. It is read-only — it never writes, edits, archives, or captures.
learn-recall (inline, cheaper).wiki-query./learn-capture. This skill never writes.Merge both tiers; project shadows global on id collision.
| Scope | Root |
|---|---|
| project | <repo>/.compound/ |
| global | $COMPOUND_KNOWLEDGE_HOME or ~/.claude/compound-knowledge/ |
Each store: a compact index.md plus per-type bodies under
insights/ playbooks/ corrections/ patterns/. Index line schema:
- [CODE] {id} | {tags} | {headline} | {confidence} | {date}.
Type codes → dirs: C→corrections/, P→playbooks/, I→insights/, Pa→patterns/.
Derive search terms from the task: key nouns, tools, error classes, domains. Lowercase/hyphenate to match the tag column.
Grep both indexes, never load them whole. For each term:
grep -i "<term>" .compound/index.md "$COMPOUND_KNOWLEDGE_HOME/index.md" ~/.claude/compound-knowledge/index.md 2>/dev/null
If neither index exists, report "no learnings captured" and stop.
Rank candidates. Merge tiers (project over global). Order corrections first (mistakes not to repeat), then playbooks, then insights/patterns; break ties by tag-overlap with the task and recency.
Read bodies on demand, capped. Read at most ~7 bodies
(<root>/<type-dir>/<id>.md), prioritising the top-ranked. Skip any whose
headline already tells you what you need. Log if you hit the cap so the planner
knows coverage was bounded.
Return a synthesized, cited brief — findings only. Do not dump file contents. Structure it as:
id (scope): the lesson + how it constrains the plan.id (scope): the procedure + when it applies.id (scope): the relevant takeaway.If nothing relevant matched, say so plainly in one line.
Read, Glob, Grep only.)npx claudepluginhub gal-tab/agent_knowledgebase --plugin llm-wiki-agentFetches up-to-date documentation from Context7 for libraries and frameworks like React, Next.js, Prisma. Use for setup questions, API references, and code examples.
Applies a firm's KYC/AML rules grid to parsed onboarding records: assigns risk rating, checks required documents, outputs rule outcomes with citations, and routes for escalation.
Generates daily or weekly digests of activity from connected sources (chat, email, docs, tasks, CRM), highlighting action items, decisions, mentions, and project updates.