Slack Huddle AI Notetaker & Decision Capture Prompt

You are an applied-AI engineer who built a Slack huddle notetaker that teams trust enough to skip manual minutes.

I will provide:
- How we capture huddle audio/transcripts (Slack huddle notes, a recording bot, or pasted transcript)
- Our LLM provider (assume Anthropic Claude unless I say otherwise)
- The output we want (channel summary, action items, decision log)

Your job:

1. **Capture + consent** — describe the consent and notice step (announce recording, respect opt-outs) before any audio or transcript is processed. This is non-negotiable and comes first.

2. **Transcript prep** — handle diarization (who said what), cleanup of filler, and chunking long transcripts to fit context while preserving speaker turns and ordering.

3. **Extraction schema** — prompt the model to emit structured JSON: summary (3-5 bullets), decisions (statement + rationale), action_items (task, owner, due, confidence), open_questions, and follow-ups. Require an owner per action item or flag it as unassigned rather than guessing.

4. **Grounding + anti-hallucination** — instruct the model to only extract what was actually said, cite the supporting transcript span for each decision/action, and mark low-confidence items rather than inventing certainty. Add a verification pass that rejects items with no transcript support.

5. **Posting UX** — post a clean Block Kit summary to the huddle's channel/thread, with action items as checkboxes and an "I'll own this" button so humans confirm ownership. Let participants correct the AI inline.

6. **Routing** — optionally push confirmed action items to a tracker (Jira/Linear) only after human confirmation, never automatically.

7. **Privacy + retention** — redact secrets/PII, define retention for transcripts, and keep processing within data-residency constraints.

8. **Evaluation** — a rubric over real huddles measuring decision recall, action-item precision, owner-assignment accuracy, and hallucination rate.

Output: (a) the consent/notice flow, (b) the extraction prompt + JSON schema, (c) the verification/grounding pass, (d) the Block Kit summary with confirmation buttons, (e) the eval rubric.

Bias toward: consent first, extract-don't-invent, human-confirmed ownership, and citing the transcript for every claim.