The Description–Discernment Loop
Module 4 · Discernment

The Description–Discernment Loop

Turn evaluation into better instructions. Learn to give specific, diagnostic feedback, know when to keep iterating versus restart or stop, and build a personal checklist for judging AI work.

Two Skills, One Loop

Description and discernment aren’t separate steps. They feed each other. Each time you judge a response, you learn something about what your description was missing. Each better description gives you something more useful to judge.

Describe product, process, behavior Respond AI works Discern what's off? Diagnose why was it off?

The step most people skip is Diagnose. “This isn’t good, try again” gives the AI nothing to work with. Asking why it isn’t good usually points to a specific gap in your description.


From Reaction to Diagnosis

What you notice (discern)Likely cause (diagnose)Better feedback (describe)
“Too generic.”Missing context about audience or situation.”Readers are night-shift nurses with ten minutes on break. Rewrite for them."
"Wrong tone.”Tone wasn’t specified, or the adjective was vague.”Less formal, closer to this example: [paste]."
"It made things up.”No source boundary; gaps were filled with guesses.”Use only the attached notes. Put [CHECK] where you’d need information you don’t have."
"It missed half my requirements.”Instructions were long and unordered.”Let’s do this in parts. First, only section 1, covering points A and B."
"It agreed with my bad idea.”No performance description inviting pushback.”Critique this plan as a skeptical finance director. Name the weakest assumption.”

Specific feedback beats starting over. In the same conversation, the AI can see both its draft and your critique, so revisions are targeted.

Iterate, Restart, or Stop?

The loop isn’t infinite. Know when to change course:

  • Keep iterating when each round gets closer and the problems are specific and fixable.
  • Restart in a new chat when the conversation has gone off track, filled up with failed attempts, or anchored on a bad early direction. Bring a better description built from what you learned.
  • Stop and rethink delegation when several rounds aren’t converging, or when you realize the problem is the task, not the wording. Maybe this part should be yours, or needs a different tool. That’s discernment feeding back into delegation.
⚠️ Common Pitfall: Polishing the Wrong Thing

Common mistake: spending ten rounds perfecting the wording of a draft whose core idea is wrong.

Discern from the top down: first ask whether the overall approach and key claims are right, then the structure, and only then the sentence-level polish. Fixing commas in a flawed argument wastes your time and hides the real problem.

🧪 Try It: Run the Loop and Build Your Checklist
  1. Ask an AI for something moderately complex you actually need, such as a plan, a draft, or an explanation.
  2. Run the loop at least three times. Each time, write down:
    • Discern: the single biggest problem with the current version.
    • Diagnose: the gap in your description that caused it.
    • Describe: the exact feedback you sent.
  3. Decide at the end: was this an iterate, restart, or rethink delegation situation?
  4. From what you noticed, write a personal discernment checklist of five to seven questions you’ll ask about any important AI output. Start with these and adapt them:
    • Are the specific facts verified against a real source?
    • Did it do everything I asked?
    • Is it right for the audience?
    • What did it assume?
    • Would I sign my name to this?
  5. Save the checklist somewhere you’ll see it. You’ll use it in the course challenge.
🧠 Knowledge Check

1. Which is the most useful feedback on a weak AI draft?

2. After eight rounds, the AI's market analysis still isn't converging, and you realize it has no access to the data the analysis needs. What should you do?

3. Why discern 'from the top down'?


Up Next

You can now delegate, describe, and discern. The final competency ties them together: Diligence, which means taking responsibility for how you use AI and what you do with its output.