The 8 ChatGPT codes that stop it from agreeing with you

ChatGPT has a default personality problem: it agrees with you. Ask whether your pricing makes sense and it finds a reason it does. Show it a plan and it tells you the plan is strong. That agreement feels like validation and costs you money, because the whole point of a thinking partner is the part where it pushes back.

The fix is not a longer prompt. It is a set of short codes you define once at the top of a chat and then use like keyboard shortcuts. "Secret codes" is a slightly dramatic name for what these are, which is a custom command vocabulary. Nothing is hidden inside ChatGPT waiting to be discovered. You are teaching it eight specific modes and then calling them by name, which is why they work in any chat, on any plan, in any AI tool.

What does a code block look like?

Paste this once at the start of a chat. Then drop any code after something you have written and it responds in that mode only.

Here are 8 codes. When I type one after something I've written, respond only in
that mode.

/ATTACK - Argue against what I just said as hard as you honestly can.
/HOLES - Find my unstated assumptions and what I left out.
/STEELMAN - Give the strongest possible case for the opposite position.
/SOWHAT - Tell me why this matters or doesn't. No throat-clearing.
/ODDS - Give me your honest confidence level, high, medium, or low, and what
would change it.
/PLAINLY - Strip the hedging. Give me the blunt friend version.
/NEXT - Name the single most important next step and say why it beats the
alternatives.
/FAILHOW - Tell me the most likely way this goes wrong before I commit.

Eight is deliberate. Fewer and you lose the range. More and you stop remembering them, which means you stop using them.

Why do codes beat writing a better prompt each time?

Three reasons, and they compound.

The first is friction. Writing "argue against this as hard as you can, find my assumptions, and tell me your confidence level" every time is enough work that you skip it exactly when you are busy, which is exactly when you most need the pushback.

The second is separation. When you ask for critique and next steps in one breath, you get a softened blend of both. One code at a time forces one mode at a time, and the answers get sharper because the model is not hedging between jobs.

The third is that you stop steering the answer. A long prompt written while you are attached to your own idea tends to smuggle in the conclusion you want. A code you defined last week, when you were calm, does not.

Which code do you use when?

The codes map to decision stages, and using the wrong one at the wrong time is why people conclude AI critique is useless.

Before you commit money or time, use /FAILHOW. It is the pre-mortem, and it surfaces the failure mode you have been walking past. Ask it before you sign anything, hire anyone, or launch.

When you are certain, use /ATTACK or /STEELMAN. Certainty is the signal that you have stopped looking. /ATTACK goes after your position. /STEELMAN builds the best version of the opposite one, which is more useful when you are about to dismiss an option too fast.

When something you wrote feels thin, use /HOLES. It names the gaps you cannot see because you wrote it: the missing number, the assumption about timing, the claim with nothing behind it.

When you are stuck between options, use /ODDS then /NEXT. /ODDS makes the model commit to a confidence level and tell you what evidence would move it, which turns a vague debate into a question you can go answer. /NEXT collapses the list into one action.

When the answer is padded, use /PLAINLY or /SOWHAT. Both cut throat-clearing. /PLAINLY strips hedging from the substance. /SOWHAT forces the model to defend why the topic matters at all, which occasionally reveals that it does not.

What is the one code that fixes vague asks?

Separate from the eight, one line fixes the most common AI mistake, which is dumping a vague request and getting a generic answer.

Before responding, ask me clarifying questions until you're 95% confident you can
complete this task well. Use only verifiable sources. Do not speculate.

The model asks you four or five questions first. Answering them takes a minute and the output improves more than any prompt rewrite would. The reason is that the questions surface context you never thought to include, which is the same problem how to give AI context solves at a larger scale.

How do you make the codes permanent?

Pasting the block into every new chat works but wears thin. Two better options depending on your setup.

On a plan with custom instructions, put the block into your personalization settings. Every chat starts with the codes loaded and you never paste again. This is the single highest-return setting in ChatGPT and most people leave it empty.

On a plan with Projects, put the block into the project instructions for whichever project holds your real work. The codes apply to every conversation inside it, alongside your business context.

Either way, the codes become part of how you use the tool rather than something you remember to set up. Reusable prompt templates covers the wider habit of writing a prompt once and using it every week.

Do these work in Claude and Gemini?

Yes. The block is plain instruction, not a ChatGPT feature, so it works in any model that follows instructions. Claude tends to give longer, more structured pushback on /HOLES and /STEELMAN. ChatGPT tends to be blunter on /PLAINLY. Gemini is strongest when the code is applied to something already sitting in your Google Workspace files.

If you are choosing between them for this kind of work, ChatGPT vs Gemini vs Claude sorts it by task rather than hype, and 8 AI mistakes beginners make covers the habits that undermine good prompts regardless of tool.

The full code block, three worked examples showing the before and after, and the additional codes are all in the prompt codes pack, free and ready to paste.