Stop role-playing your prompts: what actually lands on modern AI

You can stop telling the AI it is a “world-class expert”, choom, and you can stop bolting “think step by step” onto everything. With today’s chromed-up models, those tricks barely twitch the needle. What actually lands is boring and effective: a clear goal, the real context, the output format you want, a few constraints, and maybe a couple of examples. That is the whole rig.

The reason is simple. The newer reasoning models already grind things out internally before they answer, like a netrunner running the maze in their own head. Tools like Claude with extended thinking, Google’s Gemini thinking models, and OpenAI’s reasoning models do the “step by step” part on their own, standard-issue in these frontier rigs. So asking for it again is like reminding a chef to use the oven, a gonk move that wastes your words.

Why did role-play prompts ever work?

Back in 2023, models were weaker and more literal, cheap chrome by today’s standard. Telling one “you are a senior copywriter with 20 years of experience” nudged it toward a more careful style, because it had no strong default. Threatening it, bribing it with imaginary eddies, or chanting “think step by step” sometimes squeezed out slightly better answers.

Models have moved on. The defaults are way stronger now, so the nudge is mostly noise. You are burning eddies worth of words on theatre instead of on the thing that matters: telling the model what you actually want. That is a gonk trade.

So what works now?

Here is the short version, choom. Swap the theatre for substance.

Start with the goal

Say plainly what you want and who it is for, no corpo fluff. “Write a friendly two-line reply declining this meeting” beats “act as an executive assistant” every time. The model does not need a costume or a braindance backstory. It needs a target, like a samurai needs a mark.

Give it real context

This is the biggest lever, choom. Paste the email you are replying to. Drop in the messy notes. Tell it the audience, the tone, and anything it cannot guess. A model with the actual details will out-run any clever phrasing, no chrome required.

Name the output format

Ask for a table, three bullet points, a tweet, a 100-word summary, whatever you need. If you do not say, you get the model’s default, and then you burn a round and some eddies asking it to reshape things.

Add constraints and a couple of examples

Constraints are the ICE around the output: “no jargon”, “under 80 words”, “British spelling”. Examples are even stronger. Show one or two samples of the style you like, and the model will match them tighter than any description could. That is preem.

What about the small, fast models?

Fair caveat. Some smaller or older non-reasoning models still gain from a little structure, like a short “here is how to approach this” nudge. So if you are running a lightweight model on your phone or a cheap tier, a bit of scaffolding can help that low-grade chrome hold together. But for the capable frontier rigs most people reach for, clarity beats ceremony, and ceremony just flatlines your tokens.

The real skill is iterating

Your first prompt is a draft, not a spell. Read the answer, tell the model exactly what is off (“too formal, cut the intro, keep the second point”), and let it revise. That back-and-forth is where the real eddies come from, choom, one clean pass through the rig at a time. I dug into that habit in prompting is editing, not incantation, and broke down the parts of a solid prompt in the anatomy of a good prompt.

So next time you jack into Claude or any modern assistant, delta the role-play at the door. Tell it the goal, hand it the context, name the format, and refine from there like a fixer closing a deal. Plain and direct wins now, and that is nova.