← Exercises

the agent design game

A repeatable process for solving real problems with agents.

1. Clarify your goal

For this exercise, each of you is the CEO of your own life and work. CEOs get work done by stating goals and intentions, evaluating others’ work, and giving feedback. That is how we work with agents, too.

Start your own transcription chat

Open a new transcription chat with your own Codex agent. Turn on transcription and speak your thoughts aloud directly to the agent, using the steps below to guide you. Keep your transcript in this chat for Part 2.

10:00

  1. Find one concrete goal

    Describe one goal that really matters to your life or work and explain why it matters. It must be something a computer or agent could plausibly help solve.

  2. Find your true goal

    Is this your real goal, or is it serving some other goal? What is your true goal? In other words, what is the actual problem? Help your agent understand the “why” behind the work.

  3. Find the real blocker

    What is actually stopping you? Talk through the problems, hidden constraints, symptoms, and assumptions. Give your agent enough detail to understand what needs to change.

    • What have you already tried?
    • What, specifically, keeps failing?
    • Is this one problem or several?
    • Is the stated problem actually the problem?
    • What feels overwhelming, uncertain, uncomfortable, or scary? Is that feeling keeping you from taking the next step?
    • What information or support would help you move forward?

  4. Talk about possible solutions

    What would definitely, definitely solve your real problem? What else? What else? What might solve your problem but you’re not sure? Keep talking through your ideas until you can explain what would help eliminate your real blockers or help you achieve your goal.

  5. Make it simpler

    If you have time, try to remove steps or complexity. What is the simplest possible thing that would prove to you that this solution is possible, without weakening the end result?

Part 1 is complete when…

  • My transcript is in my Codex chat
  • I have a meaningful and clear goal, some clear understanding of my blockers, and some plausible or directional solutions

Whole-class debrief · 5 minutes

5:00
  • Who found this difficult? Why?
  • Who found this super easy? Why?

2. Teach the agent what matters

10:00

1. Ask what the agent understood

Continue in your Codex chat with the transcript from Part 1. Use GPT 6 Astra Medium to make sure you're using an intelligent enough model.

Understand my goals

Based on this transcript, what do you understand my goals and priorities to be? What would solve my problems?

GUT CHECK:

Did it describe your goals and priorities correctly? Do you feel understood, like “hell yes, that would be awesome, you get it!”?

It’s extremely important to listen to your emotions.

Yes → Save the context

Save the goal, problem, constraints, priorities, and agreed solution from this conversation in a file named GOALS.md. Put it in my personal folder. If “personal folder” is ambiguous, ask me where that is before saving. Write it so that a future agent can immediately understand what I am trying to accomplish, why it matters, what is blocking me, the solution I want, what success looks like, my constraints and preferences, and what remains uncertain. Show me the completed file and tell me exactly where you saved it.

No → Ask for an interview

I don't have a good enough problem statement or solution yet. Help me clarify my goals until we identify something that can plausibly be solved or automated with a computer. Interview me one question at a time. Help me distinguish symptoms from the actual problem, identify hidden constraints, and make the goal concrete. Once the problem is clear, brainstorm possible solutions with me. Keep refining them until I experience a “hell yes, I would love to have that solution” reaction. Do not start building until I explicitly agree that we have found that solution.

Part 2 is complete when…

  • I have a clear goal, blocker, and “hell yes” solution
  • GOALS.md exists
  • I know exactly where the file was saved

3. Design a “hell yes” proposal

15:00

Ask for a design proposal

Initial design proposal

Okay, you understand my goals and my problems well! What do you propose for an initial design? Explain how it will work. I want the simplest and easiest to understand / clearest solution that will genuinely solve my problem.

Read the proposal. Notice your reaction.

Is anything giving you a “hell no” feeling?

You do not have to accept the agent’s way of doing things. This is your computer. This is your life.

  • The scope or complexity?
  • Too technical?
  • You do not understand how it works?
  • Too much work, cost, or maintenance?
  • The wrong tool—Google Sheets when you prefer Notion?
  • It does not solve the problem you care about?

Tell the agent how you feel and why. The model cannot read your mind or know what is best for you unless you tell it.

Turn “hell no” into “hell yes”

I am having a “hell no” reaction to this proposal. Interview me briefly to determine what feels wrong. Consider scope, complexity, technical difficulty, tools, cost, privacy, maintenance, workload, parts I do not understand, and whether it solves the problem I care about. Ask one focused question at a time, then revise the proposal. Continue until I tell you it feels like “hell yes, that would be amazing.” Do not start building before I approve it.

Part 3 is complete when…

  • The proposal solves the right problem
  • I understand the approach and tools
  • The scope feels manageable
  • I have explicitly approved it: “HELL YEA, LET'S BUILD IT”

4. Keep going!!

Build the approved solution

HELL YEA, LET'S BUILD IT. Build the approved solution now. Start with the smallest version that demonstrates the solution is possible. Test your work, show concrete evidence of what is working, tell me when you need a decision or permission or when you need me to do something for you, try safe alternatives when something fails, track what remains unfinished, and do not silently expand scope. Continue building, testing, and improving for this work session.

25:00

Blocked? Do these two things.

  1. Do not quit. Stay focused for the duration of the time.
  2. Tell the agent. Say why you are blocked, stuck, confused, or unable to move forward so it can help.
I’m stuck—help me continue

Before sending: Replace the bracketed description.

I am blocked, stuck, confused, or unsure what to do next. Here is what is happening and how I feel: [DESCRIBE WHAT IS HAPPENING] Help me understand the blocker in plain language. Resolve it directly if you can. If you need something from me, ask for the smallest specific input necessary. Then tell me the single best next action and help me take it. Do not abandon the goal because this approach failed.

If your solution works, test it. Build, improve, and iterate. The first version rarely needs to be perfect; it should give you evidence that a solution is possible.

Test and improve what we built

Test the solution we built and verify that it solves the original problem in GOALS.md. Check the important user flow from beginning to end, or tell me how to test this myself. Report what works, evidence that it works, what is incomplete or broken, risks or limitations, whether it meets our definition of success, and the most valuable next improvement. Fix safe, in-scope problems you discover, then test again.

Part 4 is complete when…

  • I have tested the important flow end to end
  • I have captured evidence: a screenshot, URL, output, or observed behavior
  • I have written down what remains broken or incomplete
  • I have named the single best next action

5. Run a retrospective

A retrospective helps you continuously improve your intuition and explicit process every time you solve—or try to solve—a problem. This is how you get better at solving problems forever.

1. Talk through your retrospective with Codex

Start a new transcription chat with your own Codex agent. Speak aloud directly to the agent about what went well, what did not, and what you want to do better next time. Use the questions below to guide your reflection.

10:00
  • Do you have the same goal as when you started?
  • Did your understanding of the problem or goal change?
  • What went well?
  • What did not go well?
  • What else remains before the problem is solved?
  • Do you feel excited about continuing?
  • What did you learn?
  • What is one thing you can do better next time?
  • What is the next concrete action?

2. Run and save your retrospective

Use your transcript from the last 10 minutes and save the result as RETRO.md.
3:00
Run and save my retrospective

Based on this conversation, the work we attempted, and GOALS.md, help me run a retrospective. Ask me about anything you cannot infer. Then summarize whether I still have the same goal; how my understanding changed; what went well; what did not; what remains; whether I feel excited about continuing; what I learned; one improvement for next time; and the single best next action. Save the final retrospective as RETRO.md in my personal folder. If that location is ambiguous, ask me where it is. Show me the file and tell me exactly where you saved it.

Part 5 is complete when…

  • RETRO.md exists and I know where it is
  • I have named one improvement for next time
  • I have one concrete next action