Choose a useful first AI task

Select a task you can verify

The first useful task should teach you how the tool works in your environment. It should also produce a result that you can inspect. A small correction to a known defect often meets both conditions.

Avoid selecting the task only because the demonstration will look impressive. A broad redesign can produce many visible changes while hiding incorrect assumptions. A small task can show whether the agent reads instructions, respects scope, and reports failed checks accurately.

You do not need to select the easiest possible task. Select one where your team can recognize a correct result and explain why it is correct.

Compare candidate tasks

Consider three fictional requests in a reporting application.

The explanation task helps you inspect reasoning and evidence. The regression test adds a controlled action. The architecture task may be valuable later, but it requires a much stronger brief and review process.

For a first exercise, choose the regression test. Use invented dates and a local branch. State that production access, dependency upgrades, and unrelated refactoring are outside scope.

Write the completion condition

“Improve date handling” leaves too much interpretation. Use a specific condition: “When the input contains an invalid calendar date, return a validation error. Preserve the documented output for valid dates.”

Add examples of valid and invalid input. Identify the existing test command. Ask the agent to inspect the current behavior before it changes files. Require a short explanation of the defect and the evidence after the change.

Separate a task outcome from an activity. “The agent wrote a test” describes an activity. “The test rejects the known defect” describes evidence. A test that passes on both the correct and incorrect code does not establish the intended protection.

Observe the work process

During the exercise, record where the agent needs additional context. Check whether it reads the relevant repository instructions. Notice whether it changes files outside scope or repeats an unsuccessful approach without new evidence.

Do not correct every minor choice immediately. Let the agent complete authorized, reversible work so you can assess the result. Intervene when the next action crosses a boundary or when continued work depends on an unresolved requirement.

At completion, review the diff and execute the relevant checks. Record both the agent time and your preparation and review time. These observations help you choose the next task and improve the working instructions.

Expand one boundary at a time

If the exercise succeeds, increase one dimension of complexity. You might move from one function to a related pair of modules. You might add a documented integration. Keep the permissions and verification requirements explicit.

If the exercise fails, identify the cause before expanding scope. Missing context, an unclear requirement, an unavailable test environment, and a model limitation need different corrections. More autonomy does not resolve all four problems.

If a prototype stays in use, assign a maintenance owner. New vulnerability information can require action without a code change. See continuous vulnerability management.