Script the Work
When the work is not trivial, build the tool that does it or proves it instead of doing it by hand. A reviewer can read and rerun a script; hand edits can only be checked by redoing them.
Skip the tool only when the task is a few obvious edits visible at a glance. The bar is triviality, not repetition: a one-off still earns a tool when the tool is what makes the work checkable.
- Do the first unit by hand to learn the recipe.
- Build the smallest tool that does or proves the job, never a
framework:
- a codemod or script for edits
- a generator for repetitive files
- a query over dumped data, such as a SQLite table, for analysis
- a rerunnable check for verification
- Prove the tool. Run it on the hand-done unit and diff the result against the hand version. Fix the tool until the diff is empty.
- Make it safe to rerun. A second run over finished units changes nothing.
- Run it on every unit yourself. When the tool handles every unit in one pass, do not start subagents to hand-apply what it does.
- When the work must go to subagents, write the recipe as a skill they all read: the steps, the check each result must pass, and the files they must not touch. Keep that skill outside their write scope, so none can edit the contract.
- Commit the tool when the work outlives the session.
Applying this skill produces a file. If the diff holds no script, codemod, generator, check, or subagent skill, the skill was not applied.
Report
Reply with the tool's path, the command that reruns it, the result of the hand-versus-tool diff, and the units it changed.