We just published the RapScalYon Starter Kit, a small Claude Code plugin under the AGPL. It is version 0.1.0 and it has a lot of growing up to do. We are sharing it now anyway, because it is part of a bigger effort: making people who build business software more productive, and we would rather work on that in the open.

What it does today

Two hooks, and nothing else.

  • Verify before finishing. If a project lists its check commands in a small file, Claude cannot call a task done until they pass. A failure is handed back so the work continues instead of ending on red. It only runs when the code has changed since the last green run, so ordinary questions cost nothing.
  • Archive before compaction. When Claude's context is summarised, the full transcript is saved first, because the decisions and the rejected approaches are exactly what a summary loses. Archives are kept outside the project folder, since a transcript can contain secrets and a project can be a public repository.

You install it once and it covers every project; each project opts in with one file. Two commands install it, and the README has them.

Why we built it

Our rule is that a change is not done until the tests and the type checks pass, and the reasons for each decision are written down. For a long time that rested on an agent remembering to follow the rule. A hook does not forget. The idea came from looking at another plugin, Proof, whose own decision log and verify gate we liked. We took the two ideas, wrote our own version, and recorded in the kit's DECISIONS.md what we adopted and what we left out, and why.

How honest we are about its state

  • It has 10 automated tests for the two hooks. They pass on Linux. We have not run them on a Mac or on Windows.
  • The verify file runs whatever commands it lists. That is fine when you wrote them. Do not accept that file from a stranger's pull request without reading it.
  • Our own project enforces only its type check through the kit. Its full test suite needs a local database, so that part still depends on people and agents doing what they said.
  • It is one small piece. It does not plan work for you and it does not review your code.

Two tools we use alongside it

We recommend, and the kit's README says so, two other open-source tools that cover the sides it does not:

  • DeepWorkPlan turns multi-step work into a written plan that survives interruptions and can be resumed. The verify hook makes each task end on green.
  • codebase-memory-mcp indexes a repository into a knowledge graph, so Claude can ask for callers and call chains instead of reading whole files.

Neither is required, and the kit does not depend on them.

Where this is going

RapScalYon.surf is the framework and app builder we are developing in public, and these tools are what we reach for while doing it. We will share more of them as they are good enough to be useful to someone else, and we will keep saying plainly how finished they are. If you try the kit and it breaks, the issue tracker is the place to tell us.