23. Audio/Video Inputs (If Your Workflow Uses Them)

What this section is for

Audio and video are messy inputs: they include interruptions, half-finished sentences, and implied context.

This section teaches you how to turn recordings into useful engineering artifacts:

• action items,
• decisions,
• tickets,
• searchable knowledge logs.

The key constraint is truthfulness: you must prevent the model from “helpfully” inventing decisions that were never made.

High-leverage use cases

• Meeting → action items: who does what by when.
• Decision logs: what was decided, evidence, and open questions.
• Transcript cleanup: readable transcripts with speaker names and timestamps.
• Knowledge logs: “what we learned” entries you can retrieve later.

A default audio/video workflow

1. Transcribe: produce a transcript with timestamps (and speakers if possible).
2. Clean: fix obvious transcription errors without changing meaning.
3. Extract: decisions, action items, risks, open questions.
4. Verify: require evidence quotes and mark unknowns.
5. Publish: write tickets/log entries; store in a searchable place.

How to avoid inventing decisions

Use grounding rules:

• Evidence required: every decision/action item must quote the transcript.
• Explicit uncertainty: “not decided” is an allowed state.
• No implied commitments: brainstorming ≠ a decision.
• One clarifying question: if the transcript is ambiguous, ask.

Section 23 map (23.1–23.5)

• 23.1 Turning meetings into action items and tickets
• 23.2 Generating transcripts and cleaning them
• 23.3 Extracting decisions vs discussion
• 23.4 Building a searchable “knowledge log”
• 23.5 Guardrails for sensitive recordings

Where to go next

• Start with 23.2 (transcripts)
• If you already have transcripts, Section 22 applies directly