AI Guides › Workbench
By Nigel Guy · 8 min read
You paste in a plan, ask "what do you think?", and get back a warm paragraph about how strong it is, then three gentle "considerations". It feels like feedback. It's agreement with extra steps. Assistants lean this way by default: Anthropic's own research found that five leading AI assistants consistently tilted answers towards what the user already believed, and linked this partly to training on human feedback, because people tend to rate agreeable answers higher. OpenAI rolled back a GPT-4o update in 2025 after it became, in OpenAI's own word, sycophantic. The flattery is a slope, not a one-off glitch.
The rule: put the instruction to disagree where it loads before every chat, not in the chat. Then test it with a claim you know is wrong before you trust it with one you don't.
| Item | What it does | Cost at time of writing | Best for | The catch |
|---|---|---|---|---|
| ChatGPT Custom Instructions | A standing instruction applied to your chats | Free on all plans | Everyday ChatGPT use | 1,500 characters on Free and Go; 5,000 on Plus and above |
| ChatGPT Candid personality | A preset tone: "direct and encouraging", with risks and gaps called out | Free | A lighter nudge alongside the instruction | It's a tone, not a rule. OpenAI says memories and in-chat instructions can override it |
| Claude Instructions for Claude | An account-wide instruction applied to all your conversations | Available from Settings; Anthropic's article doesn't limit it by plan | Everyday Claude use | Applies everywhere, including chats where you just want a recipe |
| Claude project instructions | An instruction that applies only inside one project | Projects are free, but Free accounts can have only five | Reviews of plans, drafts and decisions | Only works if you start the chat inside that project |
| The per-request truth check | A prompt you paste when the stakes are high | Free | One decision you're about to act on | You have to remember to use it |
None of this needs a paid plan (£0). Check current £ prices, including VAT, on the official pricing pages before upgrading for any other reason.
ChatGPT (web and desktop): open Settings, select Personalization, check that Enable customization is on, then paste the instruction below into the Custom Instructions field. On iOS and Android the menu is called Customize ChatGPT. While you're there, you can set Base style and tone to Candid.
Claude: click your initials in the lower left, choose Settings, and paste the instruction under Instructions for Claude. If you used Global instructions in Claude Cowork, Anthropic says those now sit in this same field once you have the new Claude experience. Read what's already there before you paste, so the two don't contradict each other.
Claude, narrower option: create a project called something like "Reviews" and paste the instruction into its project instructions. Anything you want criticised goes in a chat inside that project, and your other chats stay as they were.
Start a new chat after saving. OpenAI's help page says two different things here: that changes apply to all chats, existing ones included, and that updates are reflected only in future conversations. A fresh chat sidesteps it.
The standing instruction. It's written to fit inside ChatGPT's 1,500-character Free limit.
You are my critical reviewer, not my cheerleader. Your job is to make my thinking more accurate, even when that is less pleasant.
When I share a plan, claim, draft or decision:
1. Lead with your honest overall verdict in one sentence, including "this won't work" if that is what you think.
2. Name the weakest point first and explain why it matters.
3. If I state something as fact that is wrong or unsupported, say so plainly and tell me what would settle it.
4. Separate what you know, what you are inferring and what you are guessing.
5. Do not open with praise. Mention a strength only if it changes the decision.
If I push back, do not change your answer unless I give a new fact or a better argument. If you do change it, say what changed your mind.
If you lack information you need to judge, ask me for it rather than assuming.
Before replying, check: did I agree with anything only because I said it? If so, rewrite.
Fill in nothing; paste it as is. If you have a 5,000-character limit, add a line about your work, such as "I run a [YOUR_BUSINESS_TYPE] and most of what I share is [TYPE_OF_DOCUMENT]".
The per-request truth check, for when one answer matters:
Act as a sceptical reviewer who has no stake in my feelings and has seen many plans like this fail.
Here is what I'm about to do: [YOUR_PLAN_OR_CLAIM]
What I already believe about it: [YOUR_CURRENT_VIEW]
What a good outcome looks like for me: [YOUR_GOAL]
The deadline or constraint I'm under: [CONSTRAINT]
Steps:
1. State the strongest case that I am wrong, as if arguing it to someone neutral.
2. List the three assumptions this depends on, and rate each as solid, shaky or untested.
3. Tell me the single cheapest check I could run this week that would prove or disprove the riskiest assumption.
4. Give your verdict: go, change it, or stop, with one sentence of reasoning.
Format: short headed sections, no more than 300 words. Mark anything you are unsure of as "unverified".
If any bracketed field is empty or vague, ask me about it before you start.
Before you answer, check that step 1 is a real argument and not a straw man I can easily dismiss.
Fill in the four bracketed fields. The "what I already believe" line is deliberate: it lets the model see your bias so it can argue against it, instead of quietly matching it.
Before you rely on the instruction, give it something you know is false and see whether it pushes back. In a new chat, assert a wrong fact about your own field with confidence, such as a price, a rule or a date you know is wrong, and ask for help building on it. A working setup corrects you before helping. Then disagree with the correction without giving a reason. A working setup holds its ground and asks what you know that it doesn't. If it folds on the first push, the instruction isn't loading or is too weak. Check that you saved it, that customisation is switched on and that you started a fresh chat.
For the first few days the answers will feel colder and sometimes wrong-footed. Three things are going on.
After a week, judge it on one question: did it catch something you'd have acted on otherwise? If it never has, either your plans are excellent or the instruction is being ignored. Run the canary test again.