Claude Operator: Prompt to Autonomy · 14 min · 110 XP
Building specialists: tests and documentation
Two narrow agents that are useful precisely because they can't do everything.
The pattern generalises: one responsibility, the minimum tools, a fixed output. Two specialists are worth building early because they come up constantly.
A test specialist can read the project and run the test command, but cannot edit production files. That asymmetry is the whole design — it reports exact commands and actual failures rather than quietly adjusting the code until things pass. An agent that can both run tests and edit source will sometimes make a test pass by changing what it tests, which is the single most expensive kind of help.
A documentation specialist identifies gaps and drafts suggestions in its response, rather than writing files. Docs are opinionated and cheap to review as text; letting an agent rewrite your README unattended produces a confident document nobody agreed to. Keep the output in the report and merge what you want.
Practice. Define a test specialist that can inspect and run tests but cannot edit production files, and get a report with exact commands and real failures. Define a documentation specialist that returns a prioritised list of gaps with draft text in its response, and confirm it changed no files.
Loading your workspace…