Our summary
Installing instructions can teach a coding assistant how to build a feature without changing the model that writes its code. Ewan Mak explains this distinction for the TypeSafe skill. It is a document the assistant reads, not a background program that automatically routes its next move.
The article contrasts building Jev into an application with asking Jev whether the assistant should search, read code or ask someone. That second approach needs custom control code. It must record the raw answers, use a fixed model version, handle failed calls and check permission before taking action.
A confident recommendation does not override the assistant's restrictions or approve a file change. The demonstrated cutoff needs testing on labeled tasks, and the proposed prompt was not run by Mak. Jev's low input price alone cannot show whether this extra step saves the coding assistant's usage allowance.
Key takeaways
- Treat an installed skill as reading material unless it actually includes executable tools or hooks.
- Keep the raw model answer and model version so an assistant's summary does not become the only decision record.
- Test read-only decisions first, handle service failures explicitly, and retain separate approval for actions that change or send things.