Push-Iterate-Friction Prompting
Push past average AI answers through escalating critique and focused iteration
- Difficulty
- Easy
- Time to result
- ~days to results
- Steps
- 6
- Confidence
- 92%
Push-Iterate-Friction Prompting treats the first AI response as raw material rather than a finished answer. Begin with a clear request, inspect the result, and identify the precise section that is conventional, shallow, or evasive. Apply focused criticism to that section, then increase the demand through sharper constraints, stronger comparisons, or requests for ideas outside accepted best practice. Continue the dialogue as a debate: push, test, and refine while preserving useful portions of the working document. This process can produce more differentiated material because the model receives progressively clearer signals about what the user rejects and what direction to explore. The human remains accountable for deciding when novelty becomes implausible, manipulative, unsafe, or unsupported; iteration improves exploration, but it does not validate the resulting strategy.
Origin
Extracted from Marketing Against the Grain during a live Canvas demonstration in which repeated challenges moved the model from ordinary recommendations toward increasingly unconventional ideas.
Core principles
- 01Treat the model as a thought partner rather than a vending machine
- 02Expect initial answers to regress toward conventional wisdom
- 03Challenge weak responses instead of accepting them
- 04Increase specificity and pressure with each iteration
- 05Apply friction to selected ideas without discarding useful work
- 06Keep human judgment responsible for ethics and final quality
How to run it
- 1
Set the target
State the problem, audience, and desired form of the answer. Include enough context for the model to distinguish useful novelty from random novelty.
Pro tip Define what a strong answer must do rather than relying only on adjectives such as creative or bold.
Watch out A vague initial target makes later iterations harder to evaluate.
- 2
Diagnose the average response
Review the first answer and name exactly what is generic, obvious, or missing. Separate useful foundations from weak sections.
Pro tip Compare the answer with what a competent practitioner would already suggest.
Watch out Do not reject an accurate answer merely because it sounds familiar.
- 3
Apply focused friction
Challenge the selected passage with a direct critique and request a more specific alternative. In Canvas, edit the relevant section rather than regenerating the whole document.
Pro tip Explain why the current idea fails the goal, such as being common practice or lacking a mechanism.
Watch out Aggressive wording alone does not supply useful direction.
- 4
Escalate constraints
Ask for a substantially more differentiated answer and introduce meaningful constraints, counterarguments, or unusual perspectives. Repeat only when each round clarifies the search.
Pro tip Request mechanisms, assumptions, trade-offs, and evidence instead of novelty in isolation.
Watch out Escalation can drive the model toward extreme or unethical content if boundaries are not explicit.
- 5
Debate and test
Interrogate the emerging idea as if reviewing a colleague's proposal. Ask what would make it fail, what evidence it needs, and how it compares with safer alternatives.
Pro tip Invite the model to argue against its own recommendation.
Watch out Conversational depth can create an illusion that an unverified idea is sound.
- 6
Select and validate
Keep only ideas that are relevant, ethical, feasible, and supportable. Verify important claims independently before turning the output into a plan.
Pro tip Record why the chosen idea survived the critique rounds.
Watch out The model's willingness to elaborate is not evidence that a strategy will work.
In the wild
When Canvas proposed familiar responses to declining search traffic, the hosts called the suggestions average and repeatedly requested something more unconventional. The output became more differentiated, but also drifted into manipulative virtual-persona tactics, demonstrating both the value and the danger of escalation.
→ The iterations exposed more novel directions while making the need for an explicit ethical stopping rule obvious.
A marketer receives five predictable launch concepts, marks the strongest one, and asks why it would be ignored by the target audience. After two critique rounds, the model proposes a narrower angle based on a specific customer tension. The marketer then requests counterarguments and validates the premise with interviews.
→ The process produces a differentiated, testable campaign hypothesis rather than a generic list.
Common mistakes
Accepting the first plausible answer
Initial output often reflects common patterns and may be useful but undifferentiated. Critique it before treating it as strategy.
Demanding extremity instead of quality
Requests to make an idea ever more outrageous can produce harmful or impractical material rather than insight. Escalate meaningful constraints, not shock value.
Confusing iteration with validation
A detailed idea generated after many rounds is still a hypothesis. Test its facts, ethics, and feasibility independently.
Is it for you?
Best for
Marketers, strategists, and creators using AI for exploratory thinking, research, or draft development.
Not ideal for
High-stakes decisions where persistent prompting might manufacture confidence, unsupported claims, or unethical recommendations.
From the transcript
“what you actually get with AI is the average”
“if you continue to prompt it it will give you differentiated stuff”
“you push it you push it you push it you iterate you debate you create friction”
From the episode
OpenAI Canvas Is AMAZING! The AI Tool EVERY Marketer Has Been Waiting For