One source of truth — even when AI writes the code
How an AI-native team gave engineering and product one always-current spec — even for AI-authored changes.
- Company
AI-native developer-tools startup (anonymised at the customer’s request)
- What they build
Infrastructure that gives AI coding agents persistent, shared context
- Team
A lean engineering and product team shipping daily, much of it AI-authored
- Using Specsight for
Keeping engineering and product aligned on what actually shipped, including AI-authored changes
The challenge
This team builds for AI agents, and builds with them — agents author much of the code that ships. Two problems fed each other. First, no single source of truth between engineering and product: engineers knew the code, product knew the intent, and written specs went stale on every merge. The gap between the two pictures turned into friction. Second, because so much was AI-authored, even the team couldn’t always say what had changed — a release could shift behaviour nobody had written down. There was no visibility into what the AI shipped. Knowing the current state of the product meant booking a meeting or pinging whoever last touched the code, then waiting.
Why Specsight
They wanted a source of truth that maintained itself from the code and read in plain English — one an engineer and a product person could both open without a translation step. Hand-maintained documentation was exactly what had already failed them, so anything that needed discipline to stay current was a non-starter. As an MCP-native team, they also noticed Specsight serves the living spec over the same open standard their own agents already speak.
How they use it
They connected their repositories. Specsight analysed each one and produced a living spec — every feature broken into Context / Action / Outcome scenarios describing how it actually behaves. From then on, every merge triggers a sync: Specsight re-reads the changed code, updates the affected scenarios, flags behaviour that changed, and records it in the changelog. A change report lands in everyone’s inbox — engineering and product alike — in language anyone on the team can read.
The results
Everyone sees what changed, and everyone can act on it independently — no meeting, no ping, no waiting.
Visibility into AI-authored work: a sync surfaces the behaviour change whether a person or an agent wrote it.
The update finds the team — a notification to every inbox means the source of truth is push, not pull.
Friction between engineering and product fell, because both sides work from the same current picture.
“Most of our code ships with AI in the loop, so the scary question was always: what did we just change? Now the answer’s sitting in everyone’s inbox, in plain English.”