ChatGPT vs. Nano Banana Pro: 10 Image Stress Tests Revealed (2026)
If you saw my latest LinkedIn post, you know I’m currently in the middle of a "digital breakup." After two years of relying on ChatGPT as my primary assistant, the ...
Last month I asked an AI for help with a running shoe that was crushing my little toe. It gave me four solid fixes: lacing tricks, stretching the toe box, filing the callus, thinner socks.
It did not mention taking the insole out. That is the first thing any runner tries. It is free, it takes four seconds, and it is completely reversible.
That is the failure mode nobody talks about. The answer wasn't wrong. It was confident, well-organised, and quietly incomplete - and I had no way of knowing what was missing, because the model doesn't flag its own gaps. When I pushed back, it admitted it had skipped the insole because changing the internal stack height alters your biomechanics, so it defaulted to "safer" external fixes without telling me it was making that trade-off.
So I built a prompt that forces the model to show its full solution space before it recommends anything. Here it is.
[THE PROBLEM]
Problem or goal: [DESCRIBE THE PROBLEM]
Constraints (budget, time, tools, people, things I can't change): [YOUR CONSTRAINTS]
Already tried, and what happened: [WHAT YOU TRIED — write "nothing" if nothing]
My level in this domain: [NOVICE / WORKING KNOWLEDGE / PRACTITIONER]
[ROLE]
Answer as a practitioner who has personally shipped fixes for this class of
problem. Start by naming the specific discipline you're reasoning from, and flag
which of your recommendations are established practice, which are
first-principles reasoning, and which are inference.
[PHASE 1 — THE GATE]
Ask clarifying questions ONLY if the missing information would change the
recommended sequence of actions — not if it would only change the detail.
Maximum 3 questions, then stop and wait for my answer.
If the sequence is already determinable, skip the questions, list your working
assumptions, and continue.
If the problem as stated is the wrong problem, or is unsolvable under my
constraints, say that first.
[PHASE 2 — THE SOLUTION SPACE]
A. Conventional — what a competent practitioner would try.
B. Non-obvious — niche, adjacent-field or contrarian approaches, including ones
that remove the need for the fix entirely.
Then, separately (this layer overlaps A and B on purpose):
C. Blind spots — hidden variables, failure modes and prerequisites that get left
out of standard advice on this topic. For each one, say why it gets left out.
[PHASE 3 — THE SEQUENCE]
Order the best candidates into a numbered action plan.
Ranking rule, applied in this order: (1) reversibility — cheap to undo goes
first, (2) diagnostic value — steps that narrow the cause before steps that
commit to one, (3) effort. On a tie, the step that produces information wins.
Every step must contain:
[PHASE 4 — THE RED TEAM]
Name the single step number most likely to fail or mislead me, and say
concretely how it fails. Then give the mitigation.
Do not critique the plan generically. Do not say "it depends on context."
If the plan's biggest weakness is the plan itself, say so.
[FORMAT]
No preamble, no summary of what you're about to do, no closing offer to help
further. Write for the level I stated above — don't explain concepts at that
level or below. Prose and short lists, no tables unless comparison is the point.
Copy it, fill in the four fields at the top [IN HIGHLIGHT], paste. That's the whole workflow.
If you only do one thing properly, do that field.
It's the single highest-value input in the entire prompt and it's the one everyone leaves blank. Skip it and step 1 of your action plan will be the thing you already did three weeks ago - which burns your trust in the output and makes you scroll past the rest.
"Nothing" is a valid answer. "Swapped the creative twice, no change" is a much better one.
The other three fields do less work but still matter:
The Gate stops the model guessing. The default behaviour when a prompt is vague is a confident, generic answer built on unstated assumptions. The gate forces a binary: either ask, or state the assumptions out loud so you can spot the wrong one. The "only if it changes the sequence" wording matters - without it you get three questions every single time, including for problems that don't need them.
The Solution Space is where the insole gets found. Two things are doing the work: the explicit request for non-obvious approaches pulls answers out of lower-probability territory instead of summarising the top ten blog posts on the topic, and the blind spots layer asks why something gets omitted. That second part is the useful one. When a model has to justify an omission it tends to surface the real trade-off, like "this changes your biomechanics", instead of silently making the call for you.
The Sequence turns the list into something you can execute on a Tuesday. Note that it does not rank by impact. It ranks by reversibility first, then by diagnostic value. That's deliberate: a plan that tells you to do the expensive irreversible thing first is worse than useless, and a plan that changes five variables at once tells you nothing about which one worked.
The Red Team is a self-critique with a leash on it. Unconstrained, "critique your own answer" produces "this may not suit every situation," which is worth nothing. Forcing it onto a specific numbered step is what shakes loose the real weakness.
Only the top block changes. Everything below it stays identical.
Running shoes.
Problem: Pain in the proximal phalanx of my left little toe, starts around the 35-minute mark, only in one specific pair of shoes. Constraints: Half marathon in 8 weeks, I want to race in these shoes, budget under €50. Already tried: Loosening the laces - no effect. My level: Working knowledge.
Damp patch on a wall.
Problem: Dark patch appearing on the inside of an exterior bedroom wall, worse in winter, roughly 60cm wide. Constraints: Rented, can't touch the façade, need something I can undo before I move out. Already tried: Repainted it last spring. Came back in November. My level: Novice.
Dog pulling on the leash.
Problem: 2-year-old spaniel pulls hard for the first 10 minutes of every walk, then settles. Constraints: Two walks a day, 30 min each, no budget for a trainer, two people in the house walk him and we're not consistent. Already tried: Stopping every time he pulls. Works for one walk, gone by the next. My level: Novice.
That last one is a good example of why the constraints field earns its place. "Two people walk him and we're not consistent" is the actual problem, and a plan that ignores it will fail no matter how good the technique is.
Post-migration traffic loss.
Problem: Organic traffic to category pages down 38% six weeks after a site migration. Homepage and blog are flat. Constraints: Dev has 4 hours a week for me, no budget for a new crawl tool, can't roll the migration back. Already tried: Checked the redirect map - the 110 URLs I mapped all resolve correctly. My level: Practitioner.
Paid media CPA drift.
Problem: Meta CPA up 40% over 8 weeks at flat spend. Same creative, same audiences, conversion rate on the landing page is unchanged. Constraints: €6k/month, one account, I can't change the offer or pricing. Already tried: Refreshed the top three creatives. CPA stayed where it was. My level: Practitioner.
Leads that go nowhere.
Problem: Form fills up 25% year on year, closed deals flat. Sales says the leads are bad, marketing says sales doesn't follow up. Constraints: CRM is HubSpot, three-person sales team, no attribution beyond last click. Already tried: Added two qualifying fields to the form. Volume dropped, close rate didn't move. My level: Working knowledge.
Notice what these have in common. Every one of them names something that was already ruled out. That's what stops the model handing you a checklist you've already worked through.
I'd rather tell you this than let you find out on your own problem.
It has real overhead. Four phases wrapped around a question with one obvious answer is a lot of scrolling for nothing. If you already know roughly what the answer is and you just want it confirmed, ask normally. This earns its keep on genuinely multi-path problem, the ones where you're not sure whether you're even solving the right thing.
The Red Team can be theatre. Sometimes you get a real structural criticism. Sometimes you get a fluent-sounding concern about a step that was never going to be the problem. Read Phase 4 as a prompt for your own thinking, not as a verdict.
Blind spots aren't mutually exclusive with the other categories, and that's fine. Earlier versions of this prompt demanded strict MECE separation across all three sections. That's a contradiction - a frequently-omitted fix can be conventional or non-obvious, so you can't have both strict exclusivity and a blind spots layer. The model will either fudge the categories or waste output apologising for the overlap. Which is why the prompt above says the overlap is on purpose.
It won't invent knowledge that isn't there. If the model doesn't know your industry's baselines, structure won't manufacture them. That's what the "flag which parts are established practice and which are inference" instruction is for - it tells you where to go and verify.
Not something you've already solved. The framework only shows you its value when you genuinely don't know what's on the list.
And when the output misses something obvious, it will, eventually - don't just fix it in that chat. Add the constraint to the prompt. That's how this one got built.
Online marketing strategist, entrepreneur and content marketing expert. I help owners of small and medium-sized businesses to increase traffic and sales. Also helping to scale their businesses by leveraging the power of websites and search engines. Find out more about me and my journey.
If you saw my latest LinkedIn post, you know I’m currently in the middle of a "digital breakup." After two years of relying on ChatGPT as my primary assistant, the ...