AI as a Sparring Partner, Not an Answer Key

By Better Wise

From Hats to a Sparring Partner

In the last lesson you generated friction on purpose: reverse brainstorming asked how to guarantee failure, and the Six Thinking Hats forced you to look at one idea from six angles. This lesson adds a tireless, very fast partner to that same table. The catch is that AI will happily hand you an answer before you have done any thinking, and that answer is fluent enough to feel like your own.

By the end of this lesson you can:

Tell the difference between using AI as an answer key and using it as a sparring partner.
Write prompts that pull critique, counterexamples, and unusual references instead of finished solutions.
Run a divergence protocol so the first AI output never defines your solution space.
Spot flattery, filler, and invented facts before they reach your capstone brief.

The Answer Key Trap

AI tools return confident, well-formatted, plausible text in seconds. That speed is the whole danger. The first output becomes a frame, and every idea you have afterwards turns into a variation on it. Designers call this anchoring: you think you are exploring, but you are really editing someone else's first draft.

Compare two modes on a real example. A learner is designing a bike-share scheme for a small town.

	Answer key mode	Sparring mode
Prompt	"Give me ideas for a bike-share app"	"Here is my concept in three sentences. Where does it break for a rider with no smartphone?"
Output	Map, payments, ratings, a generic feature list	A specific failure mode, a stakeholder, a question worth testing
What happens next	You pick a feature and stop thinking	You have a real problem to solve

Answer key mode ends the thinking. Sparring mode starts it.

Prompt Patterns That Ask for Friction

You get friction by asking for it explicitly. Asking for "the best idea" invites the most average idea, because that is what the model has seen most often.

Pattern	What you say	What comes back
Strongest objection	"Give the three strongest objections a skeptical funder would raise."	Risks, not praise
Counterexample hunt	"Name two real projects like this that failed, and what is different about mine."	Precedent and contrast
Red team	"Assume this failed six months from now. Write the postmortem."	Failure paths (reverse brainstorming, done by machine)
Steelman then attack	"State my idea in its strongest form, then attack that version."	A fair fight instead of a straw man
Unusual reference	"What do a lending library and a beekeeping cooperative have in common with my problem?"	Raw material for analogies
Constraint flip	"Redesign this with zero budget, and again with no app at all."	Directions you would never have tried

Tip: One pattern per message works better than five at once. You want depth on a single angle, not a survey you will skim.

Exercise 1: first two prompts. Take your capstone concept (or a placeholder: "a tool that helps neighbours share tools"). Write one strongest-objection prompt and one counterexample prompt. Run both, then underline every sentence that names a specific person, constraint, or number.

Guidance if you get stuck

Most first responses open with a compliment ("great concept!") and a general statement ("consider your users"). Skip those lines. Look for the sentence that would cost you something to answer, for example: "The scheme fails if riders only use it in good weather, so what is your winter plan?" That sentence is the sparring partner doing its job.

Protecting Your Solution Space

The prompt patterns above only help if you have something of your own to defend. Run this loop every time:

flowchart TD
    A["✍️<br/>Your 3 directions"] --> B["🗣️<br/>Critique prompt"]
    B --> C["🔀<br/>Sort: spark, generic, wrong"]
    C --> D["🛠️<br/>Revise your concept"]
    D --> E["🧪<br/>Test in the real world"]
    E --> A

Write _____ directions in your own notes before you open the tool.
Ask for divergent options, not a verdict: "Give me five directions that contradict each other."
Sort the reply into three piles: spark (worth keeping), _____ (fits any project, discard), wrong (useful as a warning).
Keep a provenance line under every idea you keep: "my observation", "AI spark on X", "analogy from Y".
Final check: can you explain your concept without opening the chat log? If not, it is still the model's idea.

Rule of thumb: the AI's first output is one sample, not the answer space.

Exercise 2: the sixth direction. Ask for five mutually contradictory directions for your capstone problem. Then write a sixth direction in one sentence that none of the five contain. If you cannot, you have been anchored, so start again from your own three directions.

Testing Whether the Critique Is Real

Two failure modes are common enough to plan for.

Flattery. Models tend to agree with whoever is talking, a behaviour called _____. Ask "rate my idea out of 10" and you will usually get an 8. Instead, ask for the strongest case against it. Then flip the framing ("I think this idea is weak, tell me why I am wrong") and compare. If the model agrees with both versions, the agreement means nothing.

Invented detail. Statistics, papers, quotes, and competitor names can be confidently false. Every number or reference must be _____ against a real source before it enters your brief.

You can also separate real critique from filler with one question: does the sentence name a specific actor, constraint, or test? "Talk to your users" fits any project. "Ask three winter commuters whether they would pay a monthly fee" is a test you can run this week.

Exercise 3: the filler filter. Copy three sentences from any AI critique. Label each one specific or filler, then rewrite one filler sentence as a testable question with a who, a when, and a number.

Guidance if you get stuck

A testable rewrite looks like: "Will five of the ten residents on my street join a tool-sharing roster within two weeks if sign-up is a single text message?" It has a group, a time window, and a pass mark. That is what your capstone test plan will need.

Field Rules From the Sparring Table

Ask for objections, not verdicts. "What is wrong with this?" beats "Is this good?"
Write your three directions first. Always. The order of operations is the whole lesson.
Reverse the framing once to check for flattery.
Verify every fact, number, and reference before it enters the brief.
Keep the chat log as a record of your thinking, never as the source of your idea.
When the critique starts repeating itself, stop prompting and go test in the real world. _____ is the signal that the tool has nothing new to give.

What You Carry Into the Capstone

AI is a sparring partner: it throws punches, you decide the strategy.
Friction comes from specific prompts: objections, counterexamples, red teams, unusual references.
Your own directions come first, so the first output stays a sample instead of a frame.
Treat agreement and invented detail as noise until you have checked them.

Next, you will turn all of this into one page: an original concept brief with a concrete test plan, so your idea leaves the chat and meets real people.

Front: Answer key mode vs sparring mode: what is the difference? Back: Answer key mode accepts a finished solution and stops thinking; sparring mode brings your own concept and asks for friction on it.

Front: What is anchoring in an AI session? Back: The first output becomes the frame, so later ideas become variations on it instead of independent directions.

Front: Which prompt yields friction: 'give me the best idea' or 'give the three strongest objections'? Back: The objections prompt; asking for the best idea tends to return the most average one.

Front: What is the steelman-then-attack prompt? Back: Ask the model to state your idea in its strongest form, then attack that version instead of a weak one.

Front: How do you protect your solution space at the start of a session? Back: Write three of your own directions before opening the tool.

Front: What three piles do you sort an AI reply into? Back: Spark (keep), generic (discard), and wrong (useful as a warning).

Front: What is sycophancy in AI models? Back: The tendency to agree with the user's framing, which makes ratings and praise unreliable as feedback.

Front: How do you test for sycophancy? Back: Flip the framing and ask the model to argue the opposite; if it agrees with both versions, the agreement is meaningless.

Front: What must happen to any statistic, study, or quote from AI? Back: It must be verified against a real source before it enters your brief.

Front: One-question test for AI filler: Back: Does the sentence name a specific actor, constraint, or test? If not, it fits any project and is filler.

Front: Why keep a provenance line under each kept idea? Back: It records whether the idea came from your observation, an analogy, or an AI spark, so authorship stays visible.

Front: What is the final ownership check for your concept? Back: You can explain it fully without opening the chat log.

Front: When should you stop prompting and start testing? Back: When the critique starts repeating itself, meaning the tool has nothing new to offer.

What to carry into your concept brief

You now have a working loop you can repeat on any idea:

You go first. Own directions and unusual references before the tool sees anything.
Ask for friction, not for answers. Rubrics, counterexamples, pre-mortems, and forced analogies give you material to react to.
Sort every reply into spark, generic, or wrong. Keep small and few, and label the source under each kept line.
Treat praise as noise. Flip the framing to test for sycophancy, and verify every fact, statistic, and citation against a real source.
Edit until the voice is yours. If a sentence could sit in any project's brief, cut it or replace it with one of your observations.
Stop prompting when the critique repeats itself. At that point the bottleneck is evidence, not ideas.

The tool is a sparring partner: it should leave you tired and sharper, not finished and dependent.

Next up is the Capstone, where you turn one of these stress-tested concepts into an original concept brief with a concrete test plan you can actually run.

Front: What is the core rule for using AI in idea work? Back: Use it for friction and critique; keep authorship of the concept yourself.

Front: What does it mean to let the first AI output define the solution space? Back: Accepting the model's framing and options as the boundaries of what you will consider, which collapses your divergent range.