kb-query

name

kb-query

description

Query, search, and generate grounded outputs from the approved live layer of an Obsidian vault. Use this skill whenever the user asks what approved notes already say, wants grounded answers, ranked local candidates, archived Q&A reuse, publishable artifacts, or static web exports from `wiki/live/`; says "query kb", "kb-search", "archive this answer", "归档这个回答", "复用旧 Q&A", "写成公众号/thread", "对外内容", "一鱼多吃", "问知识库", or "搜索批准层并回答"; or uses Chinese-LLM-Wiki terms such as `综合页`, `output/analyses`, `中文优先分析`, `原文证据摘录`, or `先读 wiki/index.md 再回答`. This is the canonical read-side skill for approved retrieval, candidate ranking, answer archiving/reuse, creator-facing publishing, and static export. Do not use for generic writing, open web research, deterministic rendering, or governance/maintenance work handled by `kb-render` or `kb-review`.

KB Query

Search, answer, and generate outputs from the approved live layer.

In Karpathy's words: "good answers can be filed back into the wiki as new pages." In this bundle, that compounding still respects draft -> review -> live. Strong query results archive cleanly, then feed later writeback or maintenance work.

When a user speaks in the simpler raw/wiki/output language from Chinese-LLM-Wiki, translate it instead of widening the truth boundary:

• 综合页 usually means a grounded reusable analysis, so start in outputs/qa/**
• output/analyses maps to archived grounded analysis, not approved truth
• 先读 wiki/index.md remains the default navigation posture

Minimal loop

1. locate the best approved live pages and briefings
2. answer or draft the artifact from wiki/live/
3. archive the result when it is substantive
4. record the smallest durable delta that should flow back through review later

Read before querying

• local AGENTS.md
• local CLAUDE.md if present
• ../obsidian-notes-karpathy/scripts/skill-contract-registry.json
• ../obsidian-notes-karpathy/references/chinese-llm-wiki-compat.md
• ../obsidian-notes-karpathy/references/archive-model.md
• ../obsidian-notes-karpathy/references/file-model.md
• ../obsidian-notes-karpathy/references/lifecycle-matrix.md
• ../obsidian-notes-karpathy/references/briefing-template.md
• ../obsidian-notes-karpathy/references/qa-template.md
• ../obsidian-notes-karpathy/references/content-output-template.md
• ../obsidian-notes-karpathy/references/activity-log-template.md
• ../obsidian-notes-karpathy/references/search-upgrades.md
• ../obsidian-notes-karpathy/references/questions-and-reflection-policy.md
• ../obsidian-notes-karpathy/references/query-writeback-lifecycle.md
• ../obsidian-notes-karpathy/references/memory-lifecycle.md
• ../obsidian-notes-karpathy/references/graph-contract.md
• ../obsidian-notes-karpathy/references/render-template.md
• ../obsidian-notes-karpathy/references/profile-contract.md
• ../obsidian-notes-karpathy/references/episode-template.md
• ../obsidian-notes-karpathy/references/web-export-template.md

Treat skill-contract-registry.json as the canonical source for required references, baseline command, and expected write surfaces.

If onkb is available, run onkb --json query scope <vault-root> first and treat its output as the deterministic baseline for live-layer retrieval boundaries.

If onkb is missing, follow the install fallback in ../obsidian-notes-karpathy/references/lifecycle-matrix.md, then rerun the same command.

Then start with:

• wiki/index.md
• wiki/live/indices/INDEX.md
• wiki/live/indices/CONCEPTS.md
• wiki/live/indices/SOURCES.md
• wiki/live/indices/TOPICS.md when present
• wiki/live/indices/QUESTIONS.md when present
• relevant wiki/briefings/{role}.md when the request maps to a role
• prior outputs/qa/

Hard boundary

• wiki/live/** is the default truth source
• raw/ and wiki/drafts/ are not retrieval truth
• MEMORY.md is collaboration context, not topic truth
• outputs/episodes/ and graph snapshots can surface candidates, but they do not outrank approved live pages
• archived outputs are reusable, but they remain artifact archive rather than approved truth

If a human explicitly asks for source inspection, you may cite raw evidence. Default synthesis should still stay on approved live coverage.

Modes

search

• use when the user wants ranked local candidates first
• output a ranked page list with short relevance notes
• do not create a new archive file unless the user asks

research

• use when the user wants a grounded reusable answer
• save substantive outputs under outputs/qa/
• this is the closest current mapping to output/analyses

publish

• use when the user wants creator-facing prose grounded in approved knowledge
• save article / thread / newsletter / talk-outline artifacts under outputs/content/
• reuse prior archived Q&A when it already contains the needed grounded synthesis

web

• use when the user wants a static browseable export from approved knowledge
• save the package under outputs/web/{slug}/index.html
• follow ../obsidian-notes-karpathy/references/web-export-template.md for the site skeleton, per-page frontmatter, slug rules, navigation structure, and search.json shape

reflect-lite

• use when the user wants a synthesis note, question-resolution memo, or gap note that should stay outside live
• archive it under outputs/qa/
• set followup_route: draft if the result exposes durable knowledge work

If the user wants slides, deterministic reports, chart briefs, or canvas outputs from approved knowledge, hand off to kb-render.

Writeback contract

Every substantive query should leave explicit follow-up metadata behind. At minimum, record:

• source_live_pages
• open_questions_touched when relevant
• writeback_candidates
• writeback_status
• followup_route

Use:

• none when the artifact is grounded and self-contained
• draft when durable knowledge should re-enter draft -> review -> live
• review when the next step is a governance or approval pass on already-prepared material

Prefer the smallest durable delta:

• extend an existing live page when the identity is already correct
• draft a new concept / entity / synthesis when the knowledge is durable but absent
• strengthen relationships or question tracking when the graph is the weak point

Checkpoint

Before archiving a substantive answer, confirm:

• the answer is grounded in approved pages
• archive reuse did not outrank live grounding
• followup_route is concrete
• any writeback_candidates are actionable instead of vague

For simple search-mode lookups, no checkpoint is needed.

Output to the user

Report:

1. which live pages and briefings were used
2. whether prior Q&A was reused
3. where the new artifact was saved
4. whether a query or publish log entry should be appended
5. whether follow-up work belongs in drafts or maintenance instead of directly mutating live