AI call 2 · Summarize the interview

runs once per participant, at interview end

Distill the private transcript into second-person bullets the participant reviews and edits. The edited bullets — not the transcript — are the only thing carried into proposal drafting.

When it runs When the interview reaches ready_for_summary (5–8 substantive answers) and the participant proceeds. Requires at least one substantive answer.

Model settings Single user message, temperature 0.1, max tokens 700. Response parsed as a JSON object.

Template variables

role_key — founder or investor
company_name, round_type — workspace metadata
transcript_text — full interview as Fiduciary: / Participant: lines

Output contract {"summary_bullets": ["You prioritize …", "You would accept …"]} — 4–8 bullets addressed to the participant.

The prompt (verbatim template)

You are summarizing a private {role_key} interview in a controlled research study about a {round_type} term sheet for {company_name}.

Create 4–8 concise bullets capturing the participant's stated priorities, constraints, uncertainty, and acceptable tradeoffs. Preserve material qualifications and numbers. Do not add facts, legal conclusions, or advice. Address the participant as "you." This summary will be shown only to the participant for review before proposal generation.

Transcript:
{transcript_text}

Return ONLY JSON:
{"summary_bullets":["You prioritize ...", "You would accept ..."]}

Server-side handling of the response

Normalization: each bullet is trimmed, given terminal punctuation, deduplicated case-insensitively, and the list is capped at 10 (12 after user edits).
User edit pass: the bullets are rendered as an editable list — the participant can add, edit, or delete any bullet before confirming. The confirmed list is what gets stored and later fed to the proposal call.
Fallback: on any failure, bullets are synthesized directly from the participant's own answers ("You prioritize: <answer>."), so the flow continues without the model.