First working slice
The smallest version of a system that does the whole job on real data, outside production, and is scored against agreed evaluations.
A first working slice is the smallest version of a system that does the whole job, from input to output, on real data. It is not a prototype with fake data and it is not the full build. It proves that the path works, and it gives the team a score it can improve.
The slice runs outside production, in an environment that can read real records but cannot change them. It is scored against the evaluation set agreed at the start, so the team and the owner see the same numbers.
In practice
I choose the slice by asking which single case, if it works from input to output, would convince the owner that the rest is possible. That case is rarely the simplest one. It is usually the one that touches the most systems, because integration is where most of the risk sits.
The slice is also where the handover starts to take shape. The runbook, the logs and the error paths are written as the slice runs, not after it. A slice that works only when the author is in the room has not yet passed.
Example
[Composite] A claims team wanted an agent to pre-fill claim forms. The first slice handled one claim type, read from the real claims system in read-only mode, and wrote its output to a review queue. It fell below the agreed threshold on the first run. The second run, after fixing the date parsing, passed, and the team moved on to the next claim type.
Common mistakes
- Building the slice on sanitised or synthetic data. The result tells you nothing about production.
- Letting the slice grow into the full system. Keep it narrow until the score is stable.
- Reporting a demo instead of a score. A demo shows that something can happen once. A score shows what happens across the set.
Related terms
Cite this definition
First working slice. Mayur Patil, mayurpatil.ai, updated 2026-11-01. https://mayurpatil.ai/glossary/first-working-slice/