Search for existing code by name and by behavior before planning a new helper, judge near-matches by their contract, point the agent at canonical examples, and review diffs for duplicates.
Measure what fills the context, shorten what loads every turn, scope rules by path and check what the scoping selects, filter output rather than information, and judge the result on completed work.
Keep durable knowledge in the repository and current state in a dated handoff note, record how each claim was verified, and have the next session re-check Git and the tests before continuing.
Start read-only, have the agent explain and trace the code, record a baseline of your own checks, give it one small verifiable task, and review against the baseline before widening the work.
An agent saying it read CLAUDE.md or AGENTS.md is not evidence. Use the records each tool keeps, gate edits on a load record with a pre-edit hook, and list the edit paths that hook cannot see.
Declare the scope before the agent starts, match the control level to the size of the change, and compare the diff with the declared scope before you keep it, because passing tests do not prove a patch stayed small.
Scope each package's conventions to that package, start the agent where your tool expects, write a precedence rule for cross-package changes, and check which rule files loaded for each changed file.
Put the project's instructions in one AGENTS.md, keep each tool's own file thin, and record separately what each tool enforces, because hooks and permissions do not travel with the instructions.
An agent that breaks your written rules is either missing them from context or skipping them after they loaded. Check which, make the rule verifiable, and move the parts a script can decide into checks you can inspect.
