The short version
Most AI training ends just before the useful part. People learn what the tools can do, try a clever prompt, and return to a week that has not changed. The missing step is structured practice on work they already own.
This playbook turns the original constraint exercise into a four-week program. It gives the team a common baseline, repeated practice, explicit safety boundaries, and a final decision about which workflows deserve to survive.
01What should AI training accomplish?#
Training should leave the team with a demonstrated workflow, not just better vocabulary. People should be able to choose a suitable task, provide context, steer the system, evaluate the result, and explain the human review required before the work is used.
- One approved, low-risk task each person can practice repeatedly
- A shared definition of good output for that task
- A short record of what the AI got wrong and how the person corrected it
- A review gate that names who remains accountable
- A before-and-after measure such as time, rework, quality, or coverage
02What should happen before the first session?#
Set the boundaries before you teach the tool. Publish the approved products, prohibited data, escalation path, and examples of low-, medium-, and high-risk work. Then ask managers to nominate recurring tasks that are useful enough to matter and safe enough to practice in a room.
| Good first task | Poor first task |
|---|---|
| Drafting a routine internal update | Making an employment decision |
| Summarizing already-approved internal material | Uploading confidential records into a personal account |
| Generating questions for a meeting | Producing unreviewed advice for a customer |
| Reformatting information the user understands | Analyzing data the user cannot independently verify |
03How does the four-week program work?#
Use the same task for four weeks. Week one builds a baseline, week two teaches direction, week three adds verification, and week four turns the result into a documented workflow. Repetition matters because it shows whether the first win was a useful pattern or a lucky output.
Week 1
Complete the task without AI, record the time and quality standard, then identify where help might be useful.
Week 2
Run a 20-minute challenge with no more than seven prompts. Define the outcome, context, constraints, and example before the first prompt.
Week 3
Repeat the task and use a review checklist: facts, omissions, tone, sensitive information, and downstream consequences.
Week 4
Document the workflow so a colleague can repeat it, compare it with the baseline, and decide whether to adopt, revise, or stop.
04How should managers run each practice session?#
Keep the session to forty-five minutes: frame the task, work in pairs, compare outputs, and debrief the failures. The manager's job is not to demonstrate perfect prompting. It is to make judgment visible and give the team permission to discuss what did not work.
| Time | Activity |
|---|---|
| 5 minutes | State the task, boundary, and quality bar |
| 20 minutes | Work in pairs using the agreed constraint |
| 10 minutes | Compare the original, AI-assisted output, and corrections |
| 10 minutes | Record one reusable pattern and one failure to avoid |
Pair a confident explorer with a domain expert, not two enthusiasts. The domain expert is usually better at spotting confident nonsense and missing context.
05How do you know the training worked?#
Look for transfer into the work and stable quality, not attendance or self-reported confidence. A successful cohort can show repeatable workflows, the corrections those workflows require, and at least one measure that improved without moving hidden review work onto a manager.
- Can someone explain why AI was appropriate for the task?
- Can a colleague reproduce the workflow from the documentation?
- Does the output meet the same quality standard as the baseline?
- Did cycle time improve after review time is included?
- Did the team stop any use case after discovering the risk or rework was too high?
Questions people ask#
- Should everyone train on the same task?
- Use the same learning sequence, not necessarily the same task. A shared task helps in the first session; role-specific recurring work is better for the remaining weeks.
- Is prompt training still useful?
- Yes, inside a real task. Teach outcome, context, constraints, examples, and iteration as parts of direction. A standalone library of prompt tricks rarely changes how work gets done.
- What if the team is skeptical?
- Give skeptics a formal role in evaluating quality and risk. They often know the workflow well enough to identify failures an enthusiastic beginner will miss.
- Can a one-hour workshop be enough?
- It can create a useful first experience. It cannot prove a workflow is repeatable. Plan at least one follow-up cycle on the same work before calling the capability established.
Sources#
- 01New OpenAI Academy courses for the next era of work
OpenAI
A progression from fundamentals to repeatable workflows, with learning treated as part of deployment and application tied to real work.
- 02OpenAI Academy
OpenAI
Current course progression covering clear instructions, context, review, reusable workflows, boundaries, and structured agent work.
- 03AI Risk Management Framework Core
NIST
Calls for role-appropriate AI risk training, defined tasks, operator proficiency, and documented human oversight.
Related
If you'd rather not do this alone
Practice on work that matters
A facilitated build day compresses the practice loop: real workflows, clear constraints, working prototypes, and a decision about what survives.
See how it worksThe timely half
What I'm seeing this week, by email.
By subscribing you agree to Substack's Terms of Use, their Privacy Policy, and their Information Collection Notice.
Playbooks stay current. The newsletter tells you what changed. Free.