Written from experience of what these tools do well and what they do confidently but badly.
The standing rules your Claude Code sessions read every time: what the project is, what must never be changed, and what counts as done here.
A second model reviewing the first one’s work, with an output contract so it returns findings instead of a summary of its own activity.
Who may run an agent against what, which changes need a human reviewer, and where the line is between assistance and autonomy.
The specific guardrails that stop an agent from “tidying up” a working subsystem it did not understand.
What makes a codebase workable for an agent: structure, naming, tests, and the context files that stop it guessing.
Establishing what the system does today, so that afterwards you can prove the refactor did not quietly change behaviour.
Describe the situation in your own words and see the procedure — and its cost — before you spend anything.
Technical expertise on subscription. Personalised technical actions, AI-ready audits and independent expert review for founders, teams and software businesses.
