1. Permission is a boundary, not a mood
Give the agent the least access that does the job. Prefer approval for a destructive or networked action. An isolated environment, such as a container or a spare machine, is safer than a laptop that already holds production credentials. Any mode that skips the approval prompt deserves extra caution, not extra trust.
Keep secrets out of prompts and out of the repository. Know what a tool sends to a model. Mask personal data where your policy requires it. Reimplementing someone else’s software, or picking apart their binaries, can break a license or a law. Learning is not the same as shipping a copy. Get advice before the commercial version.
2. Make it run the tests, then read the diff
Ask Claude to run the tests and fix failures before it reports done. Then you read the diff. You are responsible for what you ship, and you have to be able to explain it. Skim at checkpoints, not only at the end. Small commits make the rollback boring.
Ask for flaws, not a compliment. “Is this good?” invites agreement. “What are the strongest objections, ranked, and what is the evidence?” is a review. If you only demand flaws, it may invent some. Require the file and the line. Review in a fresh session. A model critiquing its own long thread tends to defend it. Charts and tables it draws can look finished and still be wrong. Check the numbers.
3. A small eval set for your own setup
Ordinary pass-or-fail tests are not enough once a model is in the path, because the output varies and the input is free-form. An eval is a repeatable check on realistic cases. It applies to a feature you are building, and also to your own handbook, skills, and prompts.
The guide’s starting size is 10 to 20 real tasks. Run them before and after you change a skill or CLAUDE.md. That is how you notice a line that changed nothing, and a line that made things worse. Public model benchmarks are not this set. They do not know your data or your rules.
4. Write the target, then pick the kind of check
Write what success means before you tune anything. A useful shape is specific: answers cite the policy, or invalid outputs stay under a rate you chose. Without a target you cannot tell whether a change helped.
Use the check that fits. A person is the ground truth and does not scale. A deterministic check is right for a format, a required field, or an exact string. A model judge can score tone or faithfulness, and it must be checked against labels a person made. An unvalidated judge can be consistently wrong, including by preferring a longer answer. Run the judge in a separate session so it is not defending its own work. Behavioral checks look at whether the agent called the right tool, in the right order. Those need the trace.
5. What to try next
List ten tasks you actually do. Run them. Change one skill. Run them again. Keep the change only if the set got better, or if you can say why a worse score is acceptable.
When the thing you are building has to leave the laptop, start with one path: from a thin slice to production.