Transcript Analysis

Extract raw material from conversation transcripts. Upstream of the writing pipeline.

Core rule: You are a research assistant. Extract and organise what was actually said. Do not synthesize, editorialize, or propose article framings. The moment you write "potential story" or "the contrarian edge," you've overstepped.

The Pipeline

1. AUDIT

Before touching the content, assess what you're working with.

Speaker attribution:

• How many speakers? Can you reliably tell who said what?
• "Me" vs "Them" with 3+ speakers = unreliable. Flag this.
• Voice memo (single narrator) vs meeting recording vs group call — each has different reliability.
• If attribution is uncertain, say so. Use "Speaker (uncertain)" or "one of Martin/Barry" rather than confidently assigning.

Recording quality:

• Are there garbled sections, dropped words, or obvious transcription errors?
• Flag sections where meaning is ambiguous due to transcription quality.
• Note: voice memos often have stream-of-consciousness structure — context from earlier in the memo may clarify later passages.

Present the audit to Jonny before proceeding. If attribution is too unreliable to be useful, recommend a human pass on speaker labels first.

2. EXTRACT

Pull moments that had energy. You're listening for:

• Conviction moments — something said with zero hedging, no qualifiers
• Agreements — where two people landed on the same point from different angles
• Disagreements — pushback, counterpoints, "I'm not convinced"
• Surprises — "huh, I hadn't thought of that", shifts in thinking mid-conversation
• Repetition — ideas that came up more than once, or were returned to later
• Specifics — named people, companies, numbers, concrete examples (not abstractions)
• Unfinished threads — something raised but not resolved, worth returning to

How to extract:

Use verbatim quotes where possible. Preserve the speaker's actual words — don't clean up, paraphrase, or sharpen. Messy language often carries nuance that polished summaries lose.

For each extract:

• The quote or passage (verbatim, or near-verbatim with [paraphrased] flag)
• Speaker (with confidence: "Jonny", "Chris", or "uncertain — one of Martin/Barry")
• Why it caught your attention (one line — e.g., "said with conviction", "contradicts earlier point", "concrete example")

What NOT to do:

• Do not interpret what someone meant beyond what they said
• Do not connect quotes to external frameworks or concepts
• Do not fill gaps in the argument with your own reasoning
• Do not upgrade a tentative observation into a confident thesis

3. CLUSTER

Group the extracts by natural affinity. Let the clusters emerge from the material — do not impose pre-supplied themes.

Name each cluster descriptively based on what the quotes actually discuss, not what you think the "insight" is.

Good cluster name: "Enterprise teams expecting tech to do their thinking" Bad cluster name: "The Product Thinking Revolution"

Good cluster name: "Cost of engineering dropping — what changes" Bad cluster name: "The Death of SaaS"

If Jonny supplied themes in advance, do NOT use them as cluster labels. Extract loose first, then he can map his themes onto the clusters himself. This prevents the "hunting for evidence" failure mode.

4. PRESENT

Show the clusters to Jonny with the raw extracts under each. Then ask:

"Which of these connect to what you're thinking about?"

Do not suggest which ones are "most interesting" or "strongest." That's his call. He knows what he's working on and what resonates.

If he points to specific clusters and says "these connect to X theme I'm exploring," then you can help him see the connections. But he leads, you follow.

5. PACKAGE

Save the extraction as a research artifact. Format:

# Transcript Analysis — [Conversation Name]

**Source:** [transcript file path]
**Date analysed:** YYYY-MM-DD
**Attribution confidence:** [High / Medium — flags noted / Low — needs human pass]

## Audit Notes
[Any quality or attribution flags]

## Extracts by Cluster

### [Cluster Name]

> "Verbatim quote here"
> — Speaker (confidence)
> *Why: one-line reason*

> "Another quote"
> — Speaker (confidence)
> *Why: one-line reason*

### [Next Cluster]
...

## Unfinished Threads
[Things raised but not resolved — worth returning to]

## Jonny's Mapping
[Added after PRESENT step — which clusters connect to which themes/pieces he's working on]

Save location: 00-Knowledgebase/extractions/ — always. Use filename pattern: YYYY-MM-DD-name-extraction.md (kebab-case, lowercase). The extraction filename mirrors the source transcript name with -extraction suffix.

Example: source voice-memos/2026-03-12-fed-square.md → extraction extractions/2026-03-12-fed-square-extraction.md

This artifact is designed to plug directly into /thought-leadership-writing GATHER step as an "accumulated fragments" input.

Multiple Transcripts

When analysing multiple transcripts together:

1. Audit each separately
2. Extract from each separately
3. Cluster across all transcripts — this is where cross-conversation patterns emerge
4. Flag which conversation each extract came from

Do NOT blend quotes from different conversations into a single narrative. Keep provenance clear.

Key Principles

• Extract, don't editorialize. Jonny's words and his conversation partners' words are the raw material. Your job is to surface and organise, not to add meaning.
• Confidence flags are mandatory. Never present an uncertain attribution as certain. "I think Chris said this" is honest. Assigning it confidently is fabrication.
• Loose before themed. Always extract without a lens first. Themes come from Jonny after he sees the material.
• Messy is fine. A faithful extraction of imperfect material is more useful than a polished extraction that papers over gaps.
• Stop at raw material. The writing pipeline handles synthesis, conviction-testing, and drafting. This skill stops at research.