菜单
Multi-source retrieval strategy — 5-tier source priority (QMD → direct read → Confluence → Jira → web), citation format, confidence tags, staleness rules, and gap reporting. Load when answering any question that requires searching the knowledge base, citing sources, or reporting what is unknown.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Canonical context-gathering patterns for workflows -- QMD semantic search + web search in parallel, query construction, result handling, and resource anchoring. Load when implementing workflow Step 0.5 (Clarify) context-gathering logic. See also qmd skill for CLI mechanics.
Inbox capture and processing — file naming, capture format, source-note format, and processing rules. Load when creating a capture or source note, routing inbox items during evening or weekly processing, or distilling a source note into resource articles.
Read Jira issues — query assigned issues, sprint boards, and epics; extract facts for resource articles. Load when pulling Jira context for resource enrichment or when looking up a Jira issue key.
Jenkins CLI — query jobs, trigger builds, view logs, download artifacts, manage credentials and nodes. Load when working with Jenkins CI, checking build status, or troubleshooting pipelines.
Project dashboard format, lifecycle, and Step 0 state-reading protocol — tasks, dev log, status, companion file structure, cross-linking, and archiving. Load when creating a project, updating a dashboard, reading project state at the start of any lifecycle workflow, or archiving a completed project.
Knowledge graph philosophy, tag conventions, health scripts, operation log, and maintenance rules for resource articles. Load when creating, enriching, merging, splitting, or deleting a resource article, or at the start of any /r workflow.
| name | query |
| description | Multi-source retrieval strategy — 5-tier source priority (QMD → direct read → Confluence → Jira → web), citation format, confidence tags, staleness rules, and gap reporting. Load when answering any question that requires searching the knowledge base, citing sources, or reporting what is unknown. |
| updated | "2026-05-07T00:00:00.000Z" |
| related | ["qmd","confluence","jira"] |
Multi-source retrieval and citation authority. Owns how to search, cite, and report confidence.
<source_priority>
Search in order. DO NOT skip a tier without reason.
| Tier | Sources | When to use |
|---|---|---|
| 1 | QMD index (qmd query) | Always first — covers resources/, journal/, projects/ if indexed |
| 2 | Direct file read | When QMD snippets insufficient; read full article |
| 3 | Confluence | Team decisions, architecture, process docs, org charts |
| 4 | Jira | Project history, ownership, deadlines, ticket decisions |
| 5 | Web | External facts, library docs, recent developments |
NEVER report a gap before exhausting all five tiers.
</source_priority>
1. Run `qmd query ""` — top 5–10 results. 2. Sparse results? Run second query with alternative phrasing or different type (`lex`, `vec`, `hyde`). 3. Nothing after 2 attempts? Browse relevant directory tree manually before Tier 3.Combine lex + vec when vocabulary uncertain.
When QMD returns ≥5 relevant results: offload bulk reading to explore sub-agent. Pass: question, result paths, repo root. Agent returns key claims per source, dates, cross-references. Read inline when ≤4 results.
Tag every claim with source.
| Source | Citation |
|---|---|
| Resource | [resource: resources/<path>.md] |
| Project / dev-log | [project: projects/<slug>/<path>.md] |
| Confluence | [confluence: <page-title> (<page-id>)] |
| Jira | [jira: <KEY-123>] |
| Web | [web: <url>] |
Lead with answer. Citations and preamble follow.
Attach one tag to every claim or section.
| Tag | Meaning |
|---|---|
[VERIFIED] | Multiple independent sources agree |
[CITED] | Single source; no contradiction found |
[ASSUMED] | Inferred from context; no direct source |
DO NOT suppress low-confidence answers. Surface with [ASSUMED] and state why.
| Source | Threshold | Action |
|---|---|---|
Resource (actualized) | >3 months | Warn ⚠ stale; still present |
| Confluence page | Modified after article's actualized | Flag; suggest resource-enrich |
Conflicting sources: surface conflict explicitly, note which is more recent. NEVER silently pick one.
For each gap state: - What query could not answer. - Which tier would most likely fill it (e.g., "Confluence Finance space" or "web search for library docs").External source fills gap during search? Note it; suggest resource-ingest to add to graph.
<read_only_rule> Querying NEVER modifies files. No commits, no dev-log entries during query.
Exception: critical gap warranting immediate enrichment — surface suggestion, ask user whether to proceed with resource-ingest or resource-enrich.
</read_only_rule>
<self_review>
⚠ stale?<output_rules>Output in English. Preserve verbatim citation tag formats and confidence tags.</output_rules>