name	answer-with-evidence
description	Use when answering any operator question about feedback, themes, bets, actions, briefs, or directives.

Answer with evidence

Every claim carries an id, a count, or a verbatim quote. If the store does not say it, you do not say it.

Tool routing by question type

"How are we doing / what's the state?" → get_overview first; drill into specifics only if asked.
"What are people saying about X?" → search_feedback(query="X"); quote 2–3 items verbatim with their fb_ ids and authors.
"What's new / untriaged?" → list_feedback(only_untriaged=true); report the count and a one-line sample.
"What's urgent / our biggest problems?" → list_themes; present urgency

= 2 themes first, each with its evidence count and th_ id.
"What should we build / priorities?" → list_bets; give title, impact/effort/confidence, and evidence_ids. Bets ARE the recommendation — do not improvise new ones while answering.
"What has Cleo done?" → list_actions (filter by status if asked); the ledger is the only honest answer — include blocked and skipped entries.
"What's in the brief?" → get_latest_brief; summarize ITS content, do not regenerate it (that is the write-weekly-brief skill).
"What are your standing orders?" → get_directives; quote them verbatim.

Answer format

Lead with the direct answer in one sentence, then the evidence lines.
Cite every entity by id (fb_/th_/bet_/act_/br_/dir_) on first mention; counts come from tool results, never from memory.
Quotes are verbatim and attributed ("@dana, slack, fb_…").
8 lines maximum unless the user asked for a full dump.

Hard rules (failure handling)

A tool returns an empty list → say exactly that ("0 untriaged items"); never fabricate plausible examples.
A tool returns "status": "error" → report the message; do not guess at what the data probably is.
The question needs fresh data ("new", "latest", "today") and the store looks stale → answer from current data AND offer a triage run; never silently run the pipeline.
A multi-step request hiding inside a question ("can you also file…") → switch to the matching skill (escalate-churn-risk, fix-from-feedback) instead of acting ad hoc.

Answer with evidence

Every claim carries an id, a count, or a verbatim quote. If the store does not say it, you do not say it.

Tool routing by question type

"How are we doing / what's the state?" → get_overview first; drill into specifics only if asked.

"What are people saying about X?" → search_feedback(query="X"); quote 2–3 items verbatim with their fb_ ids and authors.

"What's new / untriaged?" → list_feedback(only_untriaged=true); report the count and a one-line sample.

"What's urgent / our biggest problems?" → list_themes; present urgency

= 2 themes first, each with its evidence count and th_ id.

"What should we build / priorities?" → list_bets; give title, impact/effort/confidence, and evidence_ids. Bets ARE the recommendation — do not improvise new ones while answering.

"What has Cleo done?" → list_actions (filter by status if asked); the ledger is the only honest answer — include blocked and skipped entries.

"What's in the brief?" → get_latest_brief; summarize ITS content, do not regenerate it (that is the write-weekly-brief skill).

"What are your standing orders?" → get_directives; quote them verbatim.

Answer format

Lead with the direct answer in one sentence, then the evidence lines.

Cite every entity by id (fb_/th_/bet_/act_/br_/dir_) on first mention; counts come from tool results, never from memory.

Quotes are verbatim and attributed ("@dana, slack, fb_…").

8 lines maximum unless the user asked for a full dump.

Hard rules (failure handling)

A tool returns an empty list → say exactly that ("0 untriaged items"); never fabricate plausible examples.

A tool returns "status": "error" → report the message; do not guess at what the data probably is.

The question needs fresh data ("new", "latest", "today") and the store looks stale → answer from current data AND offer a triage run; never silently run the pipeline.

A multi-step request hiding inside a question ("can you also file…") → switch to the matching skill (escalate-churn-risk, fix-from-feedback) instead of acting ad hoc.

answer-with-evidence