AIYING FOR OUTCOMES
Unit 03 — Define What Success Looks Like
TIME: 15–20 minutes | LEVEL: Start | NEED: Any conversational AI | MODE: Standalone
1. OUTCOME FIRST
What will you accomplish?
By the end of this unit, you will be able to turn a desired outcome into clear signs of success before asking AI to help you achieve it.
Knowing what you want is important.
But it immediately raises another question:
How will you know when you have actually achieved it?
That is where success criteria enter.
The principle for this unit is:
An outcome gives you direction. Success criteria tell you whether you arrived.
2. PRACTICE FIRST
Define success before asking AI
Don't open AI yet.
Imagine that your outcome is:
I want to understand KNN.
It sounds perfectly reasonable.
But what does “understand” mean?
Does success mean that:
you read an explanation?
AI generated notes for you?
you can define KNN?
you can explain KNN to someone else?
you can apply it to a simple classification problem?
you can explain why changing k changes the result?
you can recognize a situation where KNN may not be appropriate?
These are very different levels of accomplishment.
Now try this yourself.
Write one real outcome:
MY OUTCOME: ______________________________
Before asking AI anything, answer:
What would I need to see, know, do, decide or change to say confidently:
“Yes. I achieved this outcome.”
Write three signs.
Success looks like:
Now make a prediction:
If AI knew these criteria before helping me, how might its assistance change?
Keep your answer.
We'll return to it.
3. DISCOVER
An outcome needs a definition of success
Suppose you tell AI:
“Help me prepare for tomorrow's presentation.”
AI can generate:
slides,
speaker notes,
questions,
examples,
visuals,
summaries,
rehearsal scripts,
and perhaps another revised version of everything it has already generated.
But when are you ready?
Without success criteria, there may be no natural stopping point.
You can keep generating.
More slides.
More polish.
More alternatives.
More prompts.
More AI output.
Yet more output does not necessarily mean more accomplishment.
Now imagine defining success before asking AI to help.
Outcome
Be ready to deliver tomorrow's presentation effectively.
Success criteria
I can explain the three main ideas without reading the slides.
I can complete the presentation within 15 minutes.
I can respond to the five most likely audience questions.
Every important factual claim has been checked.
Now AI has a much clearer role.
Instead of merely generating slides, it can:
rehearse with you,
question you,
identify weak explanations,
simulate the audience,
help check claims,
and test whether you are ready.
The definition of success has changed the nature of the collaboration.
From vague aspiration to observable accomplishment
Consider another outcome:
“I want to improve my résumé.”
What does improve mean?
AI could rewrite the résumé endlessly.
Version 1.
Version 2.
Version 12.
Each may sound increasingly polished.
But polish isn't necessarily the outcome.
Suppose the real outcome is:
Create a résumé that clearly and accurately communicates my suitability for a particular role.
Now we can define success criteria.
The résumé:
reflects important requirements of the target role;
provides evidence of relevant accomplishments;
avoids unsupported claims;
is easy for the intended reader to scan;
contains no factual errors;
accurately represents the candidate's experience.
Now “make it better” has been replaced by something much more useful:
Make it better against these criteria.
The Pattern AI view
Look underneath these examples and a reusable structure appears:
OUTCOME → SUCCESS CRITERIA → ACTION → EVIDENCE → JUDGMENT
The surface problem can change completely.
Learning
Outcome: Understand regression.
Success criterion: Solve a new basic regression problem without AI assistance.
Work
Outcome: Improve a customer-support process.
Success criterion: Reduce resolution time without sacrificing service quality.
Everyday life
Outcome: Choose an appropriate laptop.
Success criterion: The chosen machine satisfies the required workload, portability and budget constraints.
Different contexts.
Same underlying reasoning pattern.
That is the value of Pattern AI thinking:
The situation changes. The pattern can transfer.
Success criteria change AI's role
Consider this common prompt:
“Make this report better.”
AI now has a problem.
What does better mean?
Longer?
Shorter?
More persuasive?
More detailed?
More technical?
More readable?
More heavily referenced?
More suitable for senior management?
Unless you specify otherwise, AI has to infer what better means.
That gives us an important End-User AI principle:
If you don't define success, AI may implicitly define it for you.
Compare:
“Make this report better.”
with:
“Improve this report so that a non-technical executive can understand the three principal risks, see the evidence supporting each one, and identify the decision required.”
The second instruction gives AI something closer to a finish line.
Success criteria aren't merely for AI
There is an important distinction here.
We aren't defining success criteria only so AI can produce better responses.
We are defining them so we can judge whether our own effort and the AI-assisted process accomplished what mattered.
This keeps the human at both ends of the process:
At the beginning
What am I trying to accomplish?
At the end
Did I actually accomplish it?
AI operates between those two acts of human judgment.
But don't measure everything just because you can
Success criteria do not mean turning every human objective into a number.
Consider:
Have a meaningful conversation with my daughter.
Suppose we define success as:
Conversation lasts at least 30 minutes.
Thirty minutes of distracted conversation could be less meaningful than five minutes of genuine connection.
So useful success criteria can take different forms.
Quantitative
Something meaningfully measurable.
Complete the presentation within 15 minutes.
Qualitative
Something requiring informed judgment.
The recommendation clearly communicates the major trade-offs.
Binary
Something either occurred or did not.
Every cited source was independently checked.
Often the best definition of success combines these.
The objective isn't:
Measure everything.
It is:
Have enough evidence to make a sound judgment about whether the outcome was achieved.
4. POWER PROMPT
Define Success Before We Begin
Use this after identifying your outcome but before asking AI to produce the solution.
My intended outcome is: [OUTCOME].
Do not start solving or producing anything yet.
Help me define what successful accomplishment of this outcome would look like.
Ask me focused questions, one at a time, about:
what should change or become possible,
who should benefit,
what evidence would indicate success,
important constraints,
what must not be sacrificed to achieve the outcome,
and what could make an apparently good result unsuccessful in practice.
Then propose 3–5 success criteria.
For each criterion, identify whether it is:
Quantitative — meaningfully measured numerically
Qualitative — evaluated through informed judgment
Binary — achieved or not achievedKeep the criteria practical and proportionate to the importance of the outcome.
Do not assume that everything important must be numerical.
Ask me to revise and approve the criteria before proceeding.
Notice the final instruction:
AI proposes. Human judges.
5. PRACTICE EXERCISE
Make success visible
Choose the context most relevant to you.
LEARN
Suppose your outcome is:
I want to understand decision trees.
Before asking AI to teach you, define success.
Perhaps:
1. I can explain the basic idea in my own words.
2. I can interpret a simple decision tree.
3. I can explain one advantage and one limitation.
4. I can recognize a simple situation where a decision tree could be useful.
5. I can answer a new basic question without asking AI.
Now use the Power Prompt.
Compare AI's proposed criteria with yours.
Don't simply accept AI's version.
Ask:
Which criteria provide stronger evidence that I genuinely understand the concept?
WORK
Suppose your outcome is:
Improve our weekly team meeting.
What does success look like?
Perhaps:
The meeting finishes within 45 minutes.
Important decisions are recorded.
Every action has an owner.
Participants know what happens next.
Important unresolved issues are surfaced.
Now ask AI to help redesign the meeting against those criteria.
Notice what changed.
You are no longer asking:
“Give me ideas for better meetings.”
You are asking:
“Help me achieve this defined outcome.”
LIFE
Suppose your outcome is:
Plan a relaxing weekend trip.
AI might otherwise optimize for:
popular attractions,
top-rated restaurants,
maximum sightseeing,
and packed itineraries.
But your success criteria might be:
No rushed mornings.
Minimal travel between activities.
Enough unplanned time.
Stay within the agreed budget.
AI should now create a very different itinerary.
That's exactly the point.
Now transfer the pattern
Return to your own outcome from Section 2.
MY OUTCOME
SUCCESS CRITERIA
1. _________________________________
2. _________________________________
3. _________________________________
4. _________________________________
5. _________________________________
Now ask AI to attack, rather than praise, your criteria:
“Try to find ways these success criteria could mislead me into believing the outcome was achieved when it actually wasn't.”
This changes AI's role.
Instead of being your answer generator, it becomes a critical challenger.
6. JUDGE
Are you measuring what matters?
Success criteria can themselves be badly designed.
So now judge them.
Ask:
Do these criteria genuinely represent my outcome?
Am I measuring something merely because it is easy to measure?
Could I satisfy every criterion and still fail to accomplish what matters?
Could optimizing one criterion damage another?
Has AI introduced assumptions that don't fit my context?
Have I ignored something important because it is difficult to quantify?
Whose definition of success am I using?
Now perform a failure test.
Ask AI:
“Assume every success criterion has been met. Construct one plausible situation in which my real outcome has nevertheless failed.”
Read the answer critically.
If AI finds a convincing failure scenario, your criteria may be incomplete.
Revise them.
Confidence calibration
Before asking AI for its critique, rate your confidence:
How well do my criteria represent genuine success?
1 — Poorly | 2 | 3 | 4 | 5 — Very well
After the critique, revision and your own reflection, rate yourself again.
If your confidence increased, ask:
What new evidence justifies my increased confidence?
The aim isn't simply to feel more certain.
The aim is to have better reasons for being certain.
That's judgment.
7. ACCOMPLISH
Give AI a finish line
You can now bring the pieces together.
OUTCOME
What do I want to accomplish?
↓
SUCCESS CRITERIA
What would indicate success?
↓
AI ASSISTANCE
Where can AI genuinely help?
↓
HUMAN ACTION
What must I do?
↓
EVIDENCE
What actually happened?
↓
JUDGE
Was the outcome achieved?
Now use:
We have agreed on my outcome and success criteria. Help me create the simplest practical path toward achieving the outcome.
For every proposed action, explain which success criterion it supports.
Flag activities that may create impressive outputs but do not materially contribute to the outcome.
Identify what AI can assist with and what requires my own action or judgment.
Keep the plan focused on accomplishment rather than activity.
This introduces a useful filter.
Whenever AI proposes another task, ask:
Which success criterion does this serve?
If the answer is:
None.
Ask whether the activity is necessary at all.
Build an accomplishment scorecard
When the work is finished, don't merely inspect the AI output.
Evaluate the outcome.
| Success criterion | Evidence | Status |
|---|---|---|
| Criterion 1 | What actually happened? | Achieved / Partial / Not achieved |
| Criterion 2 | What actually happened? | Achieved / Partial / Not achieved |
| Criterion 3 | What actually happened? | Achieved / Partial / Not achieved |
AI can help organize the evidence.
But don't automatically let AI decide the final status.
You make the judgment.
That places the human exactly where an Outcome-First approach requires:
at the beginning, defining purpose;
and
at the end, judging accomplishment.
8. REFERENCES & EVIDENCE
What is behind this idea?
This section deliberately distinguishes established evidence from AIYING propositions. The sources below were checked against academic records and authoritative institutional sources before this version of Unit 03 was prepared.
Established evidence 1 — Goal-setting research
Edwin Locke and Gary Latham's landmark 2002 paper in American Psychologist, “Building a Practically Useful Theory of Goal Setting and Task Motivation: A 35-Year Odyssey,” summarizes 35 years of empirical research on goal-setting theory. The authors review core findings, mechanisms through which goals operate, moderators of goal effects, practical significance and directions for further research. (PubMed)
Locke & Latham — PubMed record
This gives us strong grounds for saying that clearly specified goals are an important part of purposeful human performance.
However:
Locke and Latham did not study AIYING.
Their research provides an established intellectual foundation around goal-directed activity. It does not empirically validate our specific AIYING sequence.
Established educational principle — Constructive alignment
Constructive alignment provides another useful foundation.
Monash University's teaching guidance, drawing on Biggs, describes alignment among learning outcomes, assessment tasks and learning activities. The essential idea is that what learners are expected to achieve, what they do, and how achievement is assessed should logically align. (Monash University)
Monash University — Constructive Alignment
This is especially relevant to our learning examples.
If the intended outcome is:
“The learner can apply KNN to a simple classification problem,”
then merely asking whether the learner has read an AI-generated explanation is weak evidence of that outcome.
A more aligned test would ask the learner to apply KNN to a new problem.
AIYING generalizes a similar reasoning pattern beyond education:
What is the outcome, and what evidence would legitimately demonstrate that it has been achieved?
That generalization is our proposition, not a claim made by constructive-alignment research.
Established AI guidance — Human-centred AI
UNESCO's AI Competency Framework for Students defines 12 competencies across four dimensions, one of which is a human-centred mindset. UNESCO's description emphasizes human-centred AI education, critical thinking, ethical considerations and responsible AI use. (UNESCO)
UNESCO AI Competency Framework for Students
This matters for AIYING because our approach deliberately resists turning the end user into a passive recipient of machine-generated output.
The human remains responsible for questions such as:
What matters?
What is the intended outcome?
Is AI appropriate here?
Is the result credible?
Should I act on it?
AI can participate in that reasoning.
It does not remove human agency.
Established AI risk-management guidance — Criteria can be quantitative or qualitative
NIST's AI Risk Management Framework (AI RMF 1.0) organizes its core around four functions:
GOVERN → MAP → MEASURE → MANAGE. (NIST Publications)
Within the MEASURE function, NIST states that quantitative, qualitative or mixed-method approaches can be used to analyze, assess, benchmark and monitor AI risk and related impacts. (NIST AI Resource Center)
NIST's associated AI RMF Playbook also describes AI system performance or assurance criteria being measured qualitatively or quantitatively under conditions similar to deployment settings. (NIST AI Resource Center)
NIST AI Risk Management Framework
This provides a useful supporting principle for our distinction between quantitative and qualitative criteria.
But the boundary is important:
NIST is addressing AI risk management and trustworthiness—not teaching individuals how to define personal outcomes.
We therefore use it as related evidence, not as proof of AIYING.
What is specifically the AIYING proposition?
The research above comes from different fields and serves different purposes.
AIYING brings related ideas into an End-User AI practice framework.
The following sequence is specifically our proposition:
OUTCOME → SUCCESS CRITERIA → ACTION → EVIDENCE → JUDGMENT
Likewise, these statements are AIYING propositions:
If you don't define success, AI may implicitly define it for you.
Give AI a finish line.
AI proposes. Human judges.
These should therefore not be written as though Locke and Latham, Biggs, UNESCO or NIST established the AIYING methodology.
They didn't.
Our claim is narrower:
Established research and authoritative frameworks provide relevant foundations. AIYING combines and applies related principles to the emerging problem of accomplishing outcomes with AI.
That distinction is essential to the intellectual integrity of this book.
And it demonstrates the very behaviour AIYING asks readers to develop:
Don't merely accept a plausible connection. Judge the evidence behind it.
9. FURTHER STEPS
Where can you go from here?
TRY
Take three outcomes you currently care about.
For each one, write:
OUTCOME
↓
3 SUCCESS CRITERIA
↓
EVIDENCE I WOULD ACCEPT
Then ask:
Could all three criteria be satisfied while the real outcome still failed?
If yes, improve the criteria.
EXPLORE
For one week, notice every time you ask AI to make something:
“better.”
Stop.
Define better.
Instead of:
“Make this email better.”
try:
“Make this email easier to understand, keep it under 150 words, preserve the three factual points, and make the requested action unmistakable.”
Then judge whether defining success changed the quality of your AI collaboration.
READ
Explore the four foundations used in this unit:
Goal-setting theory — how explicit goals influence purposeful activity. Locke & Latham
Constructive alignment — how outcomes, activities and evidence of achievement can be aligned. Monash University
Human-centred AI competency — why AI capability needs to coexist with human agency and critical thinking. UNESCO
AI measurement and evaluation — why evaluation can require quantitative as well as qualitative criteria. NIST AI RMF
They do not collectively prove AIYING.
They help us understand the established knowledge upon which we are building.
NEXT
Unit 04 — Decide Where AI Should Help
You now have:
OUTCOME
↓
SUCCESS CRITERIA
The next question isn't:
What can AI do?
It is:
Where in the journey toward this outcome should AI actually help?
Some parts may benefit enormously from AI.
Some may require human expertise.
Some require human judgment.
And some perhaps shouldn't involve AI at all.
Unit 04 will therefore move from defining the destination to designing the Human + AI collaboration.
