query
Skill150 starsChanged 6 days agoMITView on GitHub
Answer questions against the local Obsidian wiki — not from model memory. Use when the user runs /query, or asks in natural language about "my notes / my wiki / what I've researched / my past decisions / what do I know about X". Always read wiki/index.md FIRST and WHOLE to locate pages, then read them in full, then answer with [[wikilink]] citations. If the wiki has nothing relevant, say so explicitly before giving any general-knowledge answer. Offers to file high-value answers back into wiki/syntheses/ so explorations compound.
39 other projects use a version of this.
Adopted by
Most recently taken up by
- jakubtobiasz/second-brain-templatetheir own version, first seen Nov 15, 2026
- aibtcdev/skillstheir own version, first seen Nov 10, 2026
- hydra-db/hydradb-claude-codetheir own version, first seen Nov 8, 2026
- justmytwospence/strava-mcptheir own version, first seen Nov 8, 2026
- levi-qiao/obsidian-llm-wikitheir own version, first seen Nov 8, 2026
- prdlk/skillstheir own version, first seen Nov 1, 2026
Used with
Projects with it also use attach-db, install-duckdb and 6 more.
Other files people use alongside this one.
Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
duckdbSkillattach-db
Attach a DuckDB database file for use with /duckdb-skills:query. Explores the schema (tables, columns, row counts) and writes a SQL state file so subsequent queries can restore this session automatically via duckdb -init.
Also in 3 projects that use this
carlopiSkillinstall-duckdb
Install or update DuckDB extensions. Each argument is either a plain extension name (installs from core) or name@repo (e.g. magic@community). Pass --update to update extensions instead of installing.
Also in 3 projects that use this
carlopiSkillread-file
Read and explore any data file (CSV, JSON, Parquet, Avro, Excel, spatial, …) by filename only — resolves the path automatically. Uses DuckDB + the magic extension for format auto-detection. Installs magic if needed.
Also in 4 projects that use this
seongjaeryuSkillsave
Complete and save your current thinking to Bindly
Also in 4 projects that use this
AnthropicSkillstart
Set up your bio-research environment and explore available tools. Use when first getting oriented with the plugin, checking which literature, drug-discovery, or visualization MCP servers are connected, or surveying available analysis skills before starting a new project.
Also in 4 projects that use this
MicrosoftSkilldoctor
Diagnose an Amplifier CLI environment through the CLI doctor agent. USE WHEN a user requests CLI runtime, settings or missing-capability inspection. DO NOT USE for project bugs or other hosts.
Also in 3 projects that use this
Other versions39 projects use a version of this.
Word for word · 1
- HurricaHjz/SecondSelf★ 150100% same
Edited · 0
None found.
Wrote their own version · 38
- garrytan/gbrain★ 33ksame name
- facebookincubator/velox★ 4.3ksame name
- jason-effi-lab/karpathy-llm-wiki-vault★ 718same name
- duckdb/duckdb-skills★ 598same name
- ArtemXTech/claude-code-obsidian-starter★ 225same name
- romarayt/raytsystem-public-os★ 140same name
- StarRocks/starrocks-debug-skills★ 77same name
- MetamusicX/zissa-wiki★ 58same name
- honeydew-ai/honeydew-ai-coding-agents-plugins★ 43same name
- owenliang60-ship-it/knowledge-mgmt★ 37same name
28 more not listed
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository15 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- adoptSkill
- atticSkill
- deep-lintSkill
- delegateSkill
- export-templateSkill
- gatherSkill
- ingestSkill
- lintSkill
- outputSkill
- projectSkill
- qmd-searchSkill
- reflectSkill
- suggestSkill
- weekly-auditSkill
The file
We split this skill into the rules it gives an agent
7 rules. 0 are common, 5 are distinctive, and 2 are very specific to this project.
0
Common
Most of the 31 edited versions we read keep these.
5
Distinctive
This author's own choices. Worth a look.
2
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 17
When the answer is valuable, file it back into the wiki so knowledge compounds.
- Very specific: line 39
Read wiki/index.md whole — route mode, CLAUDE.md §5 — and locate candidate pages across every ## section.
- Distinctive: line 93
Don't over-cite: one citation at the start and end of a passage from the same page is enough.
- Very specific: line 110
The freshness line — a completion gate: Every reply that read wiki pages ends with one line, whatever the active style: Freshness: <N> …
- Distinctive: line 125
Never refuse with a bare "I don't know".
- Distinctive: line 128
When Step 4 declared missing or weak coverage, a gather proposal MAY follow the answer — never replace it — if ALL four hold: the gap is …
- Distinctive: line 143
If the answer is more than ~2 paragraphs or is comparative/analytical, ask:
“Common” is measured: kept by at least 6 in 10 of the 31 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
query — answer from the wiki, with citations
Goal
Turn a question into a deep read of the compiled wiki and a synthesized, cited answer. When the answer is valuable, file it back into the wiki so knowledge compounds.
Triggers
/query <question>- Natural language: "what do my notes say about X", "what was my past decision on Y", "search my wiki for Z"
- Mentions of: my wiki / my notes / my knowledge base / what I've researched.
Depth (see CLAUDE.md → Processing depth)
Match answer depth to the question (auto standard/concise; research is opt-in or ask-first, never
silent — this is query's rule and it is unchanged. ingest differs deliberately: it picks each
source's depth after reading and records it, because a batch of forty sources cannot carry forty asks.
One question can.)
All depths stay token-efficient — research permits more depth, never filler.
- standard (DEFAULT) — balanced, cited synthesis.
- concise — a tight, direct answer citing only the few key pages.
- research — rigorous and exhaustive: exact figures, verbatim quotes with refs, explicit treatment
of agreements/contradictions across sources, and a short "limitations / gaps" note. Higher accuracy
bar, still no filler. A filed synthesis uses academic structure +
depth: research.
Pipeline
Step 1 — Read the global index (always first)
Read wiki/index.md whole — route mode, CLAUDE.md §5 — and locate candidate pages across every
## section. Read the whole catalogue before choosing pages: lexical search or a hand-picked list
can omit pages relevant to the answer.
An index past one Read call's limit is read whole in consecutive offsets to the end, the lines read checked against wc -l wiki/index.md, never stopped at a refused or truncated first chunk.
Step 1b — A membership question goes to the lookup verb
"Is this work in the vault?" and "which of these papers do we have?" are answered by the ingest skill's
lookup verb, never by index or page greps: a phrase grep misses hyphenated page names and page names that
omit the author. The verb reads the source pages (wiki/sources/) only: raw files are not searched, so a
work captured under raw/ and not yet compiled to a source page gets no verdict from it. Give it every key
the citation carries:
python3 .claude/skills/ingest/dedup_probe.py --lookup --title "<title>" --author "<first author>" --year <year> --arxiv <arxiv-id> --doi <doi> --url <url>
For several works, write one JSON object per line (keys id, title, author, year, arxiv, doi,
url) to a file under /tmp and run python3 .claude/skills/ingest/dedup_probe.py --lookup --list <file>:
one index and one set of controls serve the whole list, and each work gets the verdict it gets alone.
Each verdict proves only what its key can:
- MATCH
<page>— a canonical identifier the citation gives is that one page's own: the work is held; cite the page. - CANDIDATE
<page>(<keys>;<why not a match>) — the work may be held. Read the page and confirm it on substance (the title, authors, year and venue in the page agree with the citation) before the answer names it; opening the page is not confirmation. A candidate printed with aconflict:stays unresolved unless the page's own text shows both identifiers belong to the one work, and the answer names the conflict. Several candidates are each read. - NO MATCH (keys tried; coverage) — report "no match under these keys" with the coverage line, which also says raw files are not searched; never "missing", "absent" or "not held". Step 4's "Nothing in the local wiki covers this" is not a membership answer.
- UNKEYED — no usable key: unresolved. Name the discarded keys and look for a better key (an identifier, a fuller title, the first author with the year); never report it as a no-match or as missing.
- Exit 2,
PROBE FAILED— no verdict at all: unresolved. Report the failure and fix it (it names every source page whose frontmatter cannot be read or parsed, and one such page stops every lookup until it is fixed); never fall back to greps and never read it as a no-match.
A NO MATCH, an UNKEYED or a failed probe never by itself justifies a missing-work claim or a gather. Before
one supports either (Step 4b's gap proposal included), verify the miss: re-run with every key the citation
gives, then search by reading (the index's Sources section for the title's distinctive words and the first
author's surname, any candidate page in full, and raw/, which the verb does not search). The instrument
for that check follows the instrument rule's own reasons (CLAUDE.md §2.2): the head verifies a short list
it can confirm item by item itself; a list whose breadth one context cannot hold goes to a verifier lane,
with that reason recorded and its two search controls run as a separate first step, reported apart from
its results. Only a verified miss is reported as not held.
Step 2 — Deep-read the targets (triage by confidence)
Open the most relevant pages in full with the read tool (or obsidian-cli). Follow ## Related
links one hop out when it helps. When candidates are many, first check their confidence cheaply
(one grep "^confidence:" <candidate files>, frontmatter only) and deep-read authoritative/high
first; pull in medium as needed; consult low/very-low only to fill gaps.
If the catalogue under-covers the question and
qmd is active (the qmd-search skill — dormant unless qmd is installed + enabled), use it as the
semantic fallback (qmd query "<q>" --json --files), confidence-rank the hits, then deep-read; otherwise
grep as usual.
Step 3 — Synthesize with citations (weighted by confidence)
- Cite every page you draw from inline as
[[Page Name]]. - For a verbatim claim, use a
> blockquote. - Don't over-cite: one citation at the start and end of a passage from the same page is enough.
- Weight by
confidence: by default resolve conflicts toward the better-grounded tier (and, if tied, the newerupdated); where the tiers are close or the stakes high, weigh the two sources' evidence rather than their badges, except a conflict between the owner's own positions, which keeps both with their dated notes (CLAUDE.md §4.7: a later date alone replaces nothing); stateauthoritative/highplainly, but attribute and hedgelow/very-low("a promotional listing claims…", "an unverified transcript suggests…"); never let avery-loworunverifiedclaim harden into an asserted fact. Flag a weak citation inline, e.g.[[X]] *(low-confidence)*: optionally forlow, always for avery-lowpage the answer or a decision rests on.
Step 3b — Freshness duty (Tier-2 flags; costs no extra reads)
Judge only the pages you have ALREADY read for this answer — never read extra pages for this duty:
- Contradiction (page vs page, or page vs clearly newer evidence just read) → surface it in the
reply and add/extend the page's
## Conflicts / Open Questionsblock (CLAUDE.md §4.4). - Staleness suspicion (old
updatedagainst newer in-wiki evidence; a superseded claim) → one line in the reply + offer the fix now; if not fixed on the spot, record it on the page as frontmatterflagged: YYYY-MM-DD <one-phrase reason>so/deep-lintreconciles it later. - Flags are annotations, not logged ops — the reconciling
deep-lintrun logs their resolution. - The freshness line — a completion gate. Every reply that read wiki pages ends with one line,
whatever the active style:
Freshness: <N> pages read · oldest [[page]] (updated YYYY-MM-DD) · <k> flagged(k= flags placed or staleness fixes made this run; usually 0). The answer is not done until the line has appeared. Naming the oldest page is the locator — the line cannot be written without looking at theupdated:dates already in hand. A reply that read no wiki pages (general-knowledge fallback) skips the line and says so.
Step 4 — Degrade gracefully (two cases, never a bare refusal)
- No coverage — if
index.mdhas nothing relevant and the question is general knowledge, say so first:Nothing in the local wiki covers this — answering from general knowledge: …then answer. Never silently pretend the wiki had the answer.
- Only low-confidence coverage — if the wiki does cover it but only at
low/very-low, still give the substantive answer from those pages, opening with an explicit warning, e.g.:⚠ Low-confidence: the only sources here are a promo listing and an auto-transcript — treat as provisional. Never refuse with a bare "I don't know".
Step 4b — Gap proposal (propose-only; CLAUDE.md §6)
When Step 4 declared missing or weak coverage, a gather proposal MAY follow the answer — never
replace it — if ALL four hold: the gap is evidenced by the search just run (the brief cites what
was searched — a zero-findings claim carries its probe, CLAUDE.md §11); load-bearing for the core
of the task (answerable-but-low coverage stays a warning, unless the pages name a stronger source
that public web capture can reach, which the proposal may then name); plausibly fixable by
public web sources (not owner-only knowledge); and not explicitly declined this session. At
most ONE proposal per reply, surfaced whatever the active style, formatted why (gap + evidence) ·
what (source kinds, rough count) · how (the literal /gather command with its page budget — never
--yes) · cost. Only the owner's explicit yes runs it — that run-spec only; gather's own gates
unchanged. Consent ledger: explicit no → dead this session (a concise gap note) ·
unaddressed → one compact reminder while the task is live, then quiet (a later re-fire on a new
task references the earlier brief concisely, never a full re-brief) · "later" → one re-offer at
the natural point. Non-interactive runs never propose — state the gap in the report.
Step 5 — File high-value answers back
If the answer is more than ~2 paragraphs or is comparative/analytical, ask:
This looks worth keeping — save it to
wiki/syntheses/?
On yes, create wiki/syntheses/<slug>.md (kebab-case) with synthesis frontmatter, a
## Sources Used section listing every cited [[page]], and register it under Syntheses in index.md.
Give it a conservative inherited confidence (a synthesis defaults to medium as agent-derived and caps at high, per §4.6; drop to
low if it rests mainly on low sources); never crystallise a low/unverified claim as asserted fact.
Report the filed page and its confidence to the user — every newly added wiki file states its level
(as ingest Step 8 does), so you can review and re-grade it. qmd: filing the page does not refresh
qmd; the qmd-search skill's freshness gate catches the index up before the next qmd search (CLAUDE.md §10).
Step 6 — Log it (only if you filed a synthesis)
A pure inline answer is NOT logged — logging is for brain-updating ops only (see CLAUDE.md §5). Only if Step 5 actually filed a synthesis, append:
## [YYYY-MM-DD] synthesis | <short question>
- **Output**: filed [[synthesis-slug]]; updated [[index.md]]Log a no-synthesis (inline-only) query only if the user explicitly asks.
Hard constraints
- Never answer substantive questions from memory — read the wiki first.
- Never silently answer when the wiki lacks coverage — declare it.
- Use
confidenceto triage, weight and hedge; on only-lowcoverage, answer with a warning, never a bare refusal. - The freshness line is a completion gate (Step 3b) — every reply that read wiki pages carries it; no done-declaration without it.
- Output in British/UK English with real
[[wikilink]]citations.
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.

