# [Write behavioural test scenarios outside the codebase as a holdout set for AI agents](https://vibedraft.app/builder-brains/tibo/write-behavioural-test-scenarios-outside)

Tibo · Building · X posts

Because an agent can 'teach to the test' if it sees unit tests, store end-to-end user-behaviour specs separately and only use them to judge whether the software actually works for a real user.

**When it applies:** AI-native development at level 3+ (agent writes most code)

Source: [Tibo: AI made 99.9% of devs slower (X long post)](https://x.com/tibo_maker/status/2026276723872936386), 24 February 2026 on x.com

Topics: [How do I test an AI feature so it doesn't break?](https://vibedraft.app/builder-brains/topics/evals), [How do I ship faster with Claude Code and Cursor?](https://vibedraft.app/builder-brains/topics/coding-agents)

Updated 2026-11-03. [Web version](https://vibedraft.app/builder-brains/tibo/write-behavioural-test-scenarios-outside)

## Same topic, other founders

- Marc Lou: [Measure AI productivity by things shipped, not by how many agents you run](https://vibedraft.app/builder-brains/marc-lou/measure-ai-productivity-by-things-shipped)
- Rob Hallam: [Route Claude Code work: cheap scout reads, top model plans and builds, fresh reviewer diffs](https://vibedraft.app/builder-brains/rob-hallam/route-claude-code-work)
- Rob Hallam: [Track multi-phase coding progress in a markdown file and /clear between phases](https://vibedraft.app/builder-brains/rob-hallam/track-multi-phase-coding-progress)
- Rob Hallam: [Run Claude Code on an €8/month VPS reachable from your phone](https://vibedraft.app/builder-brains/rob-hallam/run-claude-code-on-an-8-month)

Advice cards are rewritten summaries with a link to the original. They are not the founder's words unless shown in quotation marks.
