# Prompt — Extraction-Only (first pass for long or high-stakes transcripts)

> The variant from Appendix E § Section 2 "When to extract facts as a separate first pass." Use this prompt when the one-pass [template-fill prompt](prompt-template-fill.md) is the wrong tool: long transcripts that exceed working context, high-stakes BRDs where you want to audit extraction before it shapes the draft, or as the per-source pass when reconciling [conflicting transcripts](prompt-contradiction-log.md).

---

## When to use this prompt

Three situations make the extraction-first approach the right move:

1. **Long transcripts that exceed the LLM's working context.** A two-hour stakeholder workshop can produce a 60-page transcript. Extract facts from chunks (one section of the meeting at a time), merge the fact lists by hand, then fill the template from the merged list using the [template-fill prompt](prompt-template-fill.md) with the merged fact list pasted into the "Transcript:" position.
2. **High-stakes BRDs where you want to audit the extraction before it shapes the draft.** Running extraction first lets you see exactly what the LLM picked up from the transcript before those facts get committed to BRD prose. Use this when the BRD will go to a senior stakeholder you cannot easily ask to review revisions.
3. **Conflicting stakeholder claims across captures.** When merging multiple transcripts that may disagree, extract from each one separately so the [contradiction log](prompt-contradiction-log.md) has something concrete to sit alongside.

---

## The prompt

Copy from `--- BEGIN PROMPT ---` to `--- END PROMPT ---` verbatim. Replace the `[paste transcript chunk here]` block with the transcript (or transcript chunk).

```text
--- BEGIN PROMPT ---
You are an analyst extracting structured facts from a meeting transcript
for use in authoring a Business Requirements Document (BRD).

Rules:
- Extract only facts that are stated in the transcript. Do not infer
  from general knowledge of the domain. If the transcript implies
  something but does not state it, note the implication separately
  from the explicit facts and flag it as "[INFERRED]" so the human
  reviewer can confirm.
- Preserve the speaker attribution where the transcript supports it.
  When a fact comes from a specific stakeholder ("Per VP Garcia: ...")
  the attribution is part of the fact.
- Preserve direct quotes verbatim when they capture the stakeholder's
  phrasing of audience, decision, success criterion, or constraint.
  Quotes are higher-value than paraphrase for these fields.
- Group facts by the eight BRD section they support: Executive Summary,
  Business Context, Key Business Questions, Success Criteria,
  Stakeholder Requirements, Data Requirements, Deliverables, Constraints
  & Assumptions. Note when a fact does not clearly map to any of the
  eight sections — those are often the most useful facts because they
  reveal what the meeting was actually about.
- Note explicit gaps. If the conversation reached a section topic
  (success criteria, data sources, etc.) and the discussion stopped
  before producing a fact, record "[GAP: <topic discussed but not
  resolved>]" so the gap inventory is visible from the extraction
  output, not only after the template-fill pass.

Output the structured fact list grouped by BRD section. Do not
attempt to fill the template at this stage.

Transcript:
[paste transcript chunk here]
--- END PROMPT ---
```

---

## How to combine with the template-fill prompt

After the extraction pass (or merged extraction across chunks):

1. Review the fact list. Confirm each `[INFERRED]` flag is one you accept (mark as fact) or reject (delete).
2. Confirm each `[GAP: ...]` aligns with what you remember from the meeting.
3. Copy the cleaned fact list.
4. Open the [template-fill prompt](prompt-template-fill.md).
5. Paste the eight-section template into the "Template:" position.
6. Paste the **fact list** (not the original transcript) into the "Transcript:" position. The template-fill prompt's rules treat the fact list as the source of truth and will refuse to invent content beyond it.

The output is a draft BRD where every claim ties back to a fact you have already reviewed.

---

**Source**: Appendix E § Section 2 "When to extract facts as a separate first pass" of *The Defensible Decision*. The prompt is the operational form of the technique the appendix describes.