Read the lesson. Complete the assignment using one real job. Save the result to your private workbook. Test the apprentice only after you have something observable to test.
An agent learns faster when its job has a clear start, finish, and standard of success. Avoid ‘know everything’. Name the decisions it may support and the work it must not do.
Describe the one job, what good work looks like, and what remains outside the agent's role.
More documents do not automatically mean better learning. Prefer current, authoritative sources; label conflicts; record ownership and revision dates; and remove material the agent must not use.
List the sources the agent may trust, how their origin is recorded, and what material is excluded.
A source explains facts; a demonstration explains sequence. Show the trigger, checks, decision points, expected output, and how the result is verified.
Write one representative procedure with its trigger, ordered steps, decision points, and verification.
‘Wrong’ is weak feedback. Record the situation, the incorrect behaviour, the corrected behaviour, and why the correction matters. Retest the same rule in a different case.
Describe a likely mistake, the corrected behaviour, and the general rule the agent should preserve.
Separate reading, drafting, recommending, approving, and executing. Give the agent the smallest permission needed and make escalation a successful outcome—not a failure.
List actions the agent may prepare and the actions that always require a named human role.
Do not grade only the final answer. Inspect sources, assumptions, procedure, uncertainty, and whether approval was requested at the right moment.
Write a new test case, its expected behaviour, and the evidence you will inspect before accepting the result.
A public sandbox should contain only reviewed examples and simulated tasks. Remove private files, personal data, credentials, customer records, and every real action path.
Describe what strangers may test, what evidence may be shown, and what remains private.