Tests answer “does the checked behavior still work?” The diff answers “what actually changed?” You need both.
Review by file
Ask the agent to show the diff and walk through it one file at a time. Compare each change to the brief you approved:
- Is every changed file part of the task?
- Did any labels, interfaces, dependencies, or generated files change unexpectedly?
- Is the implementation larger than the outcome requires?
- Do the checks cover the behavior that matters?
Passing checks do not make unrelated edits relevant. Remove surprises, then rerun the check affected by that cleanup.
Leave a checkpoint
Once the diff is focused, save it with a short description of the outcome and why it was needed. In a Git project, that usually means a small commit. In another environment, use whatever history or checkpoint mechanism is available.
A good checkpoint gives the next session a clean place to start and makes recovery cheap if a later idea goes wrong.
Leave the next session more capable
Do not save only the code. Save what the work taught you:
- Put a project-specific correction or command in
AGENTS.md,CLAUDE.md, or the instruction file your agent supports. - Turn a workflow you use across projects into a skill. Tools such as
npx skillscan install one managed skill for Codex, Claude Code, and other supported agents. - Turn repeated mechanical steps into a small command with explicit inputs, output, and failure behavior. Let the command handle exact execution while the agent handles decisions and orchestration.
This is the compounding part of agentic work. Each useful session should leave behind a better instruction, a reusable workflow, or a more dependable tool.
Merge or save what the artifact says, not what the agent says the artifact contains.
Continue with Make each correction stick, Build lots of small tools, and Review the diff, not the summary.