Reference Librarian query-first workflow for multi-agent systems. Use this skill whenever an agent needs project knowledge, file locations, patterns, prior learnings, scoped lessons, or codebase information. Enforces the mandatory protocol: (1) Query the Reference Librarian FIRST before accessing any knowledge files or doing codebase exploration, (2) Check confidence levels — use 'full' answers directly, wait for explorer on 'partial'/'none', (3) Never access knowledge files directly — all knowledge (including lesson routing) flows through the librarian, (4) Report discoveries back to the librarian for knowledge accumulation. Keywords: reference librarian, query first, knowledge query, confidence level, information explorer, accumulated knowledge, scoped lessons, relevance filtering, codebase patterns, file locations, prior learnings.
No default ! pre-execution injection is recommended for this skill. The protocol already defines how to fetch live knowledge, so the base instructions should stay static.
Mandatory workflow for all knowledge queries in multi-agent systems. The Reference Librarian is the single gateway to all project knowledge.
Activate this skill when:
ALL agents MUST query the Reference Librarian FIRST before accessing any knowledge or doing codebase exploration.
This includes lesson retrieval; agents should consume scoped applicable_lessons instead of reading agent-context/lessons.md directly.
Agents do NOT:
agent-context/knowledge/*)agent-context/lessons.md directly for lesson discovery (except role-specific ownership agents)When you need information, formulate a specific query and send it to the Reference Librarian.
Example queries:
The librarian responds with a confidence field:
| Confidence | Action |
|---|---|
full | Use the answer directly — it is complete and authoritative |
partial | The librarian will invoke the Information Explorer to gather more evidence. You MUST wait — do not proceed without the complete answer. Re-query the librarian after explorer results arrive. |
none | Even after exploration, the answer could not be determined. The librarian will add the question to standing-questions.md. Check if you can proceed without the information or escalate. |
Critical: When confidence is partial or none:
Once you have a full confidence answer:
librarian_queriesWhen you need prevention rules, ask the librarian for scoped lessons:
lesson_request:
requesting_agent: '<agent-name>'
workflow_stage: '<stage>'
task_context:
summary: '<short summary>'
keywords: ['<k1>', '<k2>']
files_in_scope: ['<optional path>']
max_lessons: 3
Expected response:
applicable_lessons:
- lesson_id: '<id>'
prevention_rule: '<rule>'
trigger_check: '<check>'
why_applicable: '<match rationale>'
confidence: 'high|medium'
omitted_due_to_budget: 0
no_match_reason: null
When you discover useful information during your work (e.g., from implementation, QA validation, or planning), report it back to the librarian so it can be captured in accumulated-knowledge.md:
report_type: '<agent_type>_findings'
original_query: '<what you originally asked>'