Working with an assistant

1 · Describe a result you can check

Start with the person using the app, the action they need, and how you will recognize success. Keep the first request small.

Make a checklist for preparing a community event.
I need to add a task, mark it complete, and filter unfinished tasks.
Explain your plan before changing files.
Keep the first version usable on a phone.

1. Say what a person should be able to do

   Describe the outcome, not which technology to use. “Add a task and mark it complete” is easier to check than a list of files.

2. Ask for the plan first

   Reading the plan costs one message and saves a lot of undoing.

3. Ask it to explain the jargon

   Ask for unfamiliar terms to be explained, and for any setup you must do yourself to be listed.

2 · Review before moving on

1. Check the work is in the right project

   The proposal should belong to the project you have selected.

2. Read each approval request

   A command or permission request describes an action on your computer. Read it before you allow it.

3. Open Changes and read the diff

   Open Changes to see exactly which files were edited, how, and the assistant’s explanation.

4. Test the exact behavior you asked for

   Do the thing yourself, in the running app.

5. Save a working checkpoint

   Commit before you add the next feature, so you can come back to this state.

3 · Recover from a failed change

1. Write down what happened

   What you did, what you expected, and what happened instead. Include the visible error, with no private values.

2. Ask for a focused fix

   One problem at a time, rather than “start again”.

3. Go back if the result is worse

   Use your project’s saved version or Git history to recover before attempting another change.

The assistant and the running preview have separate lifecycles. A finished conversation does not necessarily stop the project, and restarting the app does not undo source edits.