Ask
Answer a question from the TARS corpus with citations. Trigger on "what do we know about...", "why did we...", "search my brain/tars for...", or any question about past decisions, meetings, tickets, or captured sources.From its SKILL.md
npx -y skills add jdiazromeral/tars --skill askAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
2 things to look at
- 29 days oldThe repository was created 29 days ago. New is not bad, but a brand new repository carrying a familiar-sounding name is the shape a typosquat arrives in, and there has been no time for anyone else to find a problem with it.
- 0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
SKILL.md
2.4 KB, 577 tokens by cl100k_base, as published. Nobody here has run it
Ask TARS
Vault house rules. Before acting, read
$TARS_HOME/AGENTS.mdif it exists and honor it — per-vault rules there (source allowlists, tone, privacy, output layout) override this skill's defaults on conflict.
Cheap first, escalate only when needed — a full document is the last resort, never the default (one meeting capture can be ~30k tokens).
- Hub check: if a
wiki/concepts/page matches the topic, its## Notesand## Sourceslists are a curated map of the best material. - Search with chunks:
tars search "<query>" -v --json— each hit carries its best-matching chunk verbatim (~400 tokens), which often already holds the answer. Try 2–3 phrasings if the first is thin (different vocabulary, entity names, ticket keys);--connectornarrows. - Escalate per document, not per question — only when a hit's chunk isn't
enough:
tars show <doc_id> --grep "<term>" -C 5— the lines around every match (numbered, so you can re-aim with a wider-C).tars show <doc_id> --head 40— frontmatter + opening (a meeting's Notes panel sits at the top).- Full
tars show <doc_id>only when the answer genuinely needs the whole document. Also grepwiki/notes/— promoted insights outrank raw captures when they conflict.
- Log vocabulary misses. If the first phrasing found nothing useful and a
reformulation did — the idea was in the corpus under different words — append
one line to
retrieval-misses.mdat the vault root:- <YYYY-MM-DD> · asked: "<original phrasing>" · found via: "<phrasing that worked>". Don't log corpus gaps (the answer truly isn't there) or typos — only vocabulary mismatches. This file is the trigger instrument for semantic search: it gets built when the data says lexical search fails too often, not before. - Synthesize the answer only from what the corpus supports, citing each claim with its doc id and origin (URL / path / ticket). If the corpus doesn't contain the answer, say so plainly — never fill gaps silently.
- If the answer surfaced a durable insight (a decision, a rationale, a concept worth keeping), offer to promote it — see the promote skill.
What ships with it
Read from the repository
Just SKILL.md. No reference files, no scripts.