Claude Interaction Tests

Claude Interaction Tests

**Tests for whether your AI agent actually recalls what your documentation knows.**

[](LICENSE)

⚠️ Not related to Storybook's "interaction tests". Nothing here touches your UI. The unit under test is your **documentation**, and the thing being measured is an **agent's recall of it**.

Not tests of your code. Tests of the *interaction* between an agent and the context structure you built for it — the `CLAUDE.md`, the `docs/`, the rules, the memory, the docstrings.

---

The problem

You work on a project for months with coding agents. You notice context issues. People say the model is "forgetting" or "misremembering", which leads to esoteric fixes — treat it like a person, treat it like a machine, apologise to it, threaten it.

Then you find `CLAUDE.md`, and a large context window stops looking like enough. Then you find the rest of them: `PLAN.md`, `BENCHMARK.md`, `PROJECT_STATUS.md`, `RESEARCH_FINDINGS.md`, `*_REFERENCE.md`. Then you land in **drift** — every document quietly getting ready to contradict the others, despite careful incremental updates.

So you fix it. You make `CLAUDE.md` an index. You add rules for what to do on new findings. You single-source your numbers.

**But does it work?**

That question has no answer today. You cannot unit-test whether a paragraph *fires*. You cannot grep for "did the agent recognise this had already been decided". Every existing tool checks whether documentation is **correct**; none checks whether it is **reached**.

The claim

We stopped trusting code, so we wrote tests. We now don't trust agents — and we write none.

**Test-driven development is back. Just not how you expected.** The unit under test is not a function. It is a *retrieval structure under realistic pressure*: does a fresh agent, mid-task, carrying its own vocabulary and its own momentum, recall the thing that stops it doing the wrong thing?

Its nearest