Judging by confident language
Check the source evidence before the polish.
AI Foundations 02 · the judgment workbench
How do I decide whether an AI answer is good enough?
Define what a useful result looks like, inspect the draft against its sources and decide what to keep, edit or redo. Begin with work you understand, and involve another source or reviewer when a claim exceeds your knowledge.
Ryan Wanner · recorded course lesson · February 2026
All classes and lessonsSix ways to think about the work
Every class on this page is free to read and watch. No account needed.
The historical 80/20 language illustrates responsibilities; it does not promise an 80% time saving. Judge the actual result and the full effort needed to produce it.
Can you catch the three unsupported claims?
Use the same source record for every decision. This practice is fictional and deliberately includes a flawed draft.
Fictional source record
Four visitors. Two requested floor plans. One asked for HOA documents. No offers recorded.
Your criteria: support every factual claim, retain unknowns and ask one useful next question.
Inspect the intentionally flawed draft
Copy a revision request
Revise this intentionally flawed seller update using only the fictional source notes. Return the corrected update and a keep/change/remove explanation. Do not predict offers or claim approval. SOURCE NOTES Four visitors. Two requested floor plans. One asked for HOA documents. No offers recorded. FLAWED DRAFT We saw strong demand today. Offers are expected soon, and the HOA is approved. CHECKS Every claim supported; unknowns visible; one useful next question; under 100 words; unsent draft.
Make the review repeatable
Write the win state
Name the recipient, purpose and three observable criteria. A factual seller update is a useful output; a promise of future offers is not a criterion this draft can establish.
Name who can verify what
Identify facts you can confirm from the notes and questions requiring documents or another person. Do not guess about HOA approval or legal status.
Inspect the flawed draft
Compare each sentence with the supplied showing record. Find the claims of strong demand, expected offers and HOA approval. None is supported by these notes.
Choose keep, edit or redo
Make one decision and explain the defect it addresses. A specific revision instruction is more useful than asking the model to make everything better.
Compare the revised result
Use the same source notes and criteria. Keep accurate information, remove invented claims and ask the useful unresolved question.
Retain the review checklist
Save the successful criteria and one lesson from the comparison. Stop when they pass; repeated prompting is not an achievement by itself.
Three decisions in the recording
what's the win state?
keep, edit, redo
I would build my strengths first.
Treating 80/20 as a measured saving
The recording uses a division-of-work illustration; it does not measure your time savings.
Iterating without a criterion
Obtain a missing fact instead of repeatedly rewriting an unsupported sentence.
Keep a record of your decision
Create a free student portal account to download this class's files.
Download the practice, answer key and review logRead the complete workbench and answer key
# Showing-update judgment workbench Current companion to AI Foundations 02 · November 8, 2026. Fictional teaching case. ## Source record Four visitors. Two requested floor plans. One asked about HOA documents. No offers recorded. ## Intentionally flawed draft We saw strong demand today. Offers are expected soon, and the HOA is approved. ## Your decision Win-state criteria: Who can verify each fact: Keep / edit / redo: Reason: Revision 1: Result against the same criteria: Optional revision 2, only if needed: ## Answer key Reject strong demand as an unsupported interpretation. Remove offers expected: none were recorded. Remove HOA approved: the notes only record a document request. One defensible revision: Four visitors attended. Two requested floor plans and one asked for HOA documents. No offers are recorded in these notes. Can we confirm the current HOA documents are available for that request? Save the review rule, not a fictional outcome. No message was sent.
Questions about judging the result
Editorial answers for practicing this lesson.
Do I have to memorize six models?
No. Write the desired result, inspect one draft and decide what to change.
Why start with work I know?
You need a credible way to judge whether the output is correct and useful.
How many revisions should I run?
Enough to meet the stated criteria. If a fact is absent, get the fact rather than polish the guess.
Does 80/20 promise an 80% time saving?
No. It is an illustration of responsibilities, not a measured or guaranteed outcome.
What if I cannot judge a technical claim?
Mark it unresolved and use the appropriate source or qualified reviewer.
Selected teaching passages
Selected caption excerpts from the historical recording. These are deliberately non-contiguous passages, not the full transcript. Current practice and corrections are written separately above; the original recording is unchanged.
3:27 · Define the win
what's the win state? What are you looking for exactly?
5:50 · Keep, edit or redo
So again, you read the output, you orient to it, you decide keep, edit, redo, and then you act. So you send it or iterate it, and then you you kind of execute that plan.
10:40 · Start with your strengths
If it were me, I would build my strengths first. And the reason for that is one, I understand the process a lot better. And two, I'm learning AI, right? And so if I understand the process really well, I don't have to learn that process. I have to learn how AI works within that process.
17:37 · Choose work you understand
So kind of your action plan here is write down your circle of competence. What are the three areas that you know really well? What are the processes that you have that are working effectively? What are the things that potentially you might want to change?
Choose the next part of your workflow
- From Asking Questions to Delegating AI Work
Define a bounded job before asking AI to produce it.
- Revise and Save Useful AI Prompts
Keep a reversible version of a request that works.
- Build a Context Card That Stays Useful
Keep relevant facts current and resolve conflicting notes.
Judgment in the weekly work
Apply your review criteria to a recurring task.
The exercise makes review concrete. The Agent Constraint Accelerator is a next step to consider when you want help applying that judgment to your own operating constraint.