Complete your first AI workflow

Now put the course into practice. Choose one small behavior from a project you can run and test locally.

A good capstone is server-side display-name validation:

• accept names from 1 to 50 characters
• reject empty and 51-character names
• preserve the current database schema and success response
• do not rely on browser validation
• keep the change inside the profile update flow

Use your own task if it has similarly clear boundaries.

1. Write the task contract

Record the outcome, current behavior, scope, constraints, and acceptance criteria. Another developer should be able to decide whether the task is complete without reading your mind.

Do not prescribe unknown implementation details. Ask the agent to inspect the system first.

2. Create an inspection note

Have the agent find the entry point, data flow, tests, and contract to preserve. Require file paths and separate observed facts from inferences.

Review the note before editing. Your goal is to catch one wrong assumption while it is still cheap.

3. Work through checkpoints

Use one behavior and one check per checkpoint:

1. Add a failing boundary test.
2. Implement server validation.
3. Run focused and nearby tests.
4. Review the diff and scope.
5. Run the production build if the change affects integration.

Approve only the tools and targets needed for the current step. Stop if the agent needs a credential, destructive action, or scope expansion you did not authorize.

4. Create a verification receipt

Save this artifact with your result:

Outcome:
Files inspected:
Files changed:
Checks run:
Failure or abuse case tried:
Data and permission decisions:
Unsupported assumption caught:
What remains unverified:
One workflow improvement for next time:

Include the provider and exact model identifier when repeatability matters. Save the final prompt and the most useful context, but do not save secrets or private customer data.

Evaluate the workflow

Success is not the AI finished quickly. Success means the requested behavior is supported by evidence, the change stayed in scope, sensitive data remained protected, and uncertainty is explicit.

If you caught and corrected a model mistake, that is a successful workflow. Verification did its job.