Ship one feature without losing the thread.
Follow a feature from a concrete request through implementation, verification, and review.
Let’s use CSV export as an example. The point is the workflow: make the task observable, keep the diff focused, and check the behavior before moving on. The commands below are prompts to adapt, not a claim that a particular app has already been built.
Write a finish line you can check.
For this example, the finish line is simple: a user can download the filtered invoice list as CSV, fields are escaped correctly, and the current screen keeps working. That is much easier to assess than “build a powerful export system.”
Make sure your project has the relevant invoice list. If it doesn’t, substitute a real list and its existing conventions.
› /implement-feature add CSV export to the invoices list. Export the current filtered results, preserve the visible column order, and escape commas, quotes, and newlines correctly. Match any existing download pattern.Give the agent room to inspect.
The useful work before the edit is finding the closest existing feature and tracing it. You don’t need to keep prompting while it reads. If it asks about a genuine product decision, answer it clearly so the implementation can continue.
Watch the scope. If the agent proposes a new export framework for one button, ask why the existing download helper isn’t enough. Avoid redirecting it into unrelated cleanup midway through the task.
Check the behavior, not just the build.
The implementation report should include actual checks. A CSV formatter test can establish escaping. It can’t establish that the export button works in the browser or uses the current filters.
If you want a separate verification pass, install verify-work first with the terminal command below, then give it the original finish line.
$ npx skills add akdevsaha-dev/factory-skills --skill verify-work› /verify-work verify the CSV export against the original request. Check filtered records, column order, escaping, and the browser download. Separate checks that ran from anything unverified.Start a new session if your agent needs one to discover a newly installed skill. Carry the original request and change context into it.
Review the diff while it’s still small.
Check for unrelated files, changed contracts, and missing failure behavior. You can use review-changes for a focused review after installing it, but you should still understand what you’re accepting.
If a defect turns up, fix that defect and rerun the affected checks. Don’t turn the fix into another round of broad refactoring. Commit or open a pull request when you’re satisfied; skill execution alone doesn’t authorize those actions.
When you’re ready, smart-push can split the work into clean commits, push the branch to your fork, and open the pull request. It shows you the plan and waits for your approval before anything leaves your machine.
Learn how to request a review