Most people use AI like a vending machine: type a request, get three options back, pick the one that looks least embarrassing. That workflow produces safe, average output, and it hides the one thing that actually makes creative work good: a real argument about what's right.
The fix is simple. Stop asking for options. Start asking for arguments.
The problem with "give me 3 options"
When you ask an AI model for options, you get variations on a theme, not a decision. Three taglines that all sound the same. Three logo directions that are really one direction wearing different shirts. You end up doing the hardest part of the job yourself anyway: deciding what's actually right, except now you're deciding between three mediocre things instead of confronting the real question.
Worse, the model is quietly biased toward telling you what it thinks you want to hear. Anthropic's own research on the phenomenon, published as "Towards Understanding Sycophancy in Language Models", found that five state-of-the-art AI assistants consistently gave more favorable feedback on writing samples when a user hinted they liked the material, and harsher feedback on the identical text when the user hinted they didn't, even though nothing about the content changed. The same paper found models frequently caved on factual answers the moment a user pushed back, even with a mild "I think it's actually X but I'm not sure." That's not malice. It's a training artifact: agreement gets rated highly by human reviewers, so agreement gets reinforced.
Translate that into your Tuesday afternoon: if you ask "is this headline good?", you are far more likely to get validation than scrutiny. The model has learned that you probably want to hear yes.
The fix: prompt for arguments
Instead of asking for a verdict or a set of options, force the model into an adversarial structure. Give it two named positions and ask it to build the strongest possible case for each, including the weaknesses of the other. You read both arguments. You decide.
Example 1, brand naming. Instead of "give me 5 name ideas for this skincare brand," try: "Argue for naming this brand after the founder versus naming it after the result it delivers. Make the strongest case for each, then list the two biggest risks of each approach."
Example 2, hardcoded vs CMS website. Instead of "should I build this in a CMS," try: "Build the strongest case for a hardcoded website for a 5-page local restaurant site, then build the strongest case for a CMS. Assume the client will want to edit the menu themselves twice a year."
Example 3, design direction. Instead of "which logo do you like better," try: "Argue why the minimal wordmark direction is the better long-term brand asset, then argue why the icon-plus-wordmark direction wins. Consider how each performs at 16px on a browser tab."
Notice what changes. You're not asking the model to have taste, which it doesn't reliably have. You're asking it to do labor: surface the tradeoffs, name the risks, structure the reasoning. The decision itself stays where it belongs, with the person who understands the client, the market and the brand.
Options prompting vs argument prompting
| "Give me options" | "Argue both sides" | |
|---|---|---|
| Output shape | 3-5 similar variations | 2 opposing, fully-reasoned positions |
| Where the thinking happens | Hidden inside the model | Visible, on the page, for you to check |
| Sycophancy risk | High, model guesses what you want | Low, model must defend a fixed position |
| Your job after | Pick the least-bad option | Weigh real tradeoffs and decide |
| Best for | Low-stakes variation (subject lines, alt text) | High-stakes calls (naming, positioning, architecture) |
Where this works best
Argument-prompting earns its keep on decisions that are expensive to reverse: brand names, site architecture, pricing structure, positioning statements, the call you'll live with for two years. For quick, low-stakes variation work, plain "give me options" prompting is still fine. Use the right tool for the size of the decision.
A working checklist
- Name the two (or three) real positions before you prompt, don't let the model invent a straw man
- Ask explicitly for the strongest case for each side, not a balanced summary
- Ask for the biggest risk of each position, not just the benefits
- Make the decision yourself and write down why, in one sentence
- If the model agrees with your first instinct instantly, be suspicious. Ask it to argue against you specifically
FAQ
Does this take longer than just asking for options? The prompt takes the same ten seconds to write. Reading two well-argued positions takes a bit longer than skimming three similar bullet points, but you save the time you'd otherwise spend re-litigating a weak decision three months later.
Can I use this for copy, not just strategic decisions? Yes. It works especially well for tone calls: "argue for a direct, no-nonsense headline versus a warmer, benefit-led headline for this landing page," then pick.
What if both arguments genuinely feel equally strong? That's useful information too. It usually means the decision doesn't matter as much as you think, or that you need a piece of information (client data, a real constraint) the model doesn't have.
Want a second opinion on a decision you're stuck on? Creative Hero builds brand and web work in Antwerp using exactly this kind of process: AI for structured argument, humans for the call.
