From expert workflow to AI-ready spec — without weeks of shadowing.
Product teams building AI assistants and automations spend weeks shadowing domain experts to capture how they actually make decisions. What gets written up still needs translation before engineering or an AI agent can build from it.
Unsaid compresses that into a single session. The expert records themselves doing the work. You probe the recording for anything unclear. The output is a structured spec where every rule is tied to a quoted excerpt and timestamp — ready to hand to your team or an AI agent.
What the output looks like
spec.md · excerpt# Support triage agent
## Rules
### R3: Escalate repeat-reporter bugs same-day
If 3+ distinct customers report a bug within 24h, escalate to
engineering — even if severity would normally be "low."
> "If three people hit it in a day it's never really a corner case.
> I just escalate." — Interview, 06:42
## Test cases
| Input | Action |
|------------------------------|----------|
| 1 reporter, sev=high | escalate |
| 4 reporters in 18h, sev=low | escalate |
| 2 reporters in 30h, sev=low | triage |
## Open questions
- Does "distinct customers" count repeat tickets from the same
account? Clarify with SM-Lead before shipping.
Not a template — generated from an expert's own words, with every rule tied to a quoted excerpt and timestamp.