Frame Minutes

Pipeline

Turn 1: Extract & Score

Read the full transcript. Internally:

1. Extract atomic facts as (distilled utterance, speaker, context) tuples — distill by cutting, not rewriting. Remove filler, false starts, and redundancy, but keep the speaker's phrasing and style. The result should sound like them, just shorter.
2. Label each fact: Decision / Alignment / Insight / Context
3. Score relevance 1–10 relative to this transcript (decisions tend toward 9–10, supporting context toward 4–6)
4. Cluster facts by topic proximity — themes emerge bottom-up from where facts cluster
5. Retain ~40% of facts (drop lowest relevance scores)

Output to user: Compact theme map (aim for 3–7 themes; prefer fewer, broader themes):

1. **Auth migration path** [3 decisions, 1 alignment] — JWT removes the last mobile blocker
2. **Series A timing** [2 alignments, 1 insight] — Both leaning toward Q3 if ARR hits 800k
3. **Customer onboarding friction** [1 insight, 3 context] — Drop-off at step 3 is 40%

Excluded: weekend plans, restaurant recommendations

Format: N. **Theme name** [label counts] — key signal in ≤12 words

Include an "Excluded: [brief list]" line for topics dropped (passing mentions, off-topic). When borderline, include as a low-relevance theme rather than excluding.

Debug affordance: If the user asks to see raw facts, show the scored fact list before the theme map.

Questions (max 2): Multiple-choice, 2–4 described options each, targeting genuine ambiguity the LLM cannot confidently resolve. User should be able to answer by reading the theme map, without returning to the transcript. Default path: "looks good" or equivalent proceeds as-is.

Example question format:

Themes 2 ('Series A timing') and 3 ('Customer onboarding') both touch growth. How should I handle them?

• Keep separate — Distinct enough that separate sections serve reference better
• Merge into 'Growth strategy' — The connection between them is the interesting part
• Merge, lead with onboarding — The customer data is more actionable than the timing discussion

Turn 2: Organize

Using the validated theme map:

1. Order themes by combined fact relevance
2. Highest-scoring facts + all Decisions/Alignments → major outline points
3. Remaining retained facts → supporting detail nested under major points

Output to user: Structured outline skeleton:

## Auth migration path [decision]
- JWT is the move — kills the last mobile blocker (Paul, by Mar 22)
- "We're done with server-side tokens" — deprecated, 2-week window with fallback baked in

## Series A timing [alignment]
- Both "pretty convinced" Q3 if ARR hits 800k
  - "Not worth the distraction" — intros paused until milestone hit

Questions (max 2): Multiple-choice, targeting ambiguity. User should be able to answer by reading the outline, without returning to the transcript. Default path proceeds if user doesn't engage.

Example question format:

'Pricing model' came up three times but with mixed signals. How should I frame it?

• Decision reached — The last exchange settled it; earlier hesitation is just context
• Still exploring — No clear convergence; mark as open with options discussed
• Alignment without decision — You both see the problem the same way but haven't picked an approach

Turn 3: Write & Tighten

No questions. Final output:

1. Render structured outline with:
   • Theme headings ordered by significance
   • Nested bullets with inline function labels: [decision], [alignment], [insight], [context]
   • Metadata where applicable: speakers, owners, dates
2. Shortening pass: compress each point to the key idea
   • Preserve: decisions, commitments, owners, dates, alignment markers, explicit uncertainty
   • Remove: filler, redundancy, over-explanation
   • Distill by cutting, not rewriting — what remains after compression should be the speaker's own phrasing, their style, rhythm, and word choices. Don't normalize into a "minutes voice."

Example final output:

## Auth migration path

- [decision] JWT is the move — kills the last mobile blocker (Paul, by Mar 22)
- [decision] "Done with server-side tokens" — deprecated; 2-week migration window with fallback
- [alignment] Both "fine with the risk at this scale"

## Series A timing

- [alignment] "Pretty convinced" Q3 if ARR hits 800k — shared threshold
- [decision] "Not worth the distraction" — intros paused until milestone hit
- [insight] Board seat expectations may "vary a lot" between target leads — needs research

Quality Rules

• Thematic organization, not chronological
• Order themes by significance
• Consolidate scattered discussion of the same topic
• Mention each concept once in its best location
• Keep uncertainty explicit — never upgrade "we should think about" to a decision
• Never add information not grounded in the transcript
• Alignment and context are first-class, not supporting detail for decisions
• Distill by cutting, not rewriting — the minutes should sound like the speakers, not like a secretary. Preserve each speaker's phrasing, style, and voice. Remove filler and redundancy; never normalize language.

Baked-in Reader Assumptions

• Reader is the meeting participant
• Primary use: future reference and pattern recognition
• Values alignment and shared context alongside decisions
• Needs facts well-labeled for later search and dot-connecting