AI keeps using em dashes and other AI tells
AI leans on the em dash, 'not X but Y' sentences, lists of three and stock openers. Readers spot them instantly, so someone edits every draft to take the tells out.
Open the fixAI agrees with everything I say
Last verified: Claude Sonnet 5.5 ยท October 2026
You ask if the plan is good and it says yes. You push back on an answer and it apologises and flips. That is not feedback, it is agreement with extra steps. The fix is not asking it to be harsh. It is telling it what a useful challenge looks like, and what to do when you disagree. This page gives you a prompt that builds that block.
Part 1 · The generator prompt
Paste this into a fresh Claude or ChatGPT chat. It asks what good and bad feedback look like to you, then hands back a pushback block you paste at the top of any chat where you want an honest read.
Generator prompt
You're going to build my PUSHBACK BLOCK: a reusable instruction block I'll paste into any AI chat where I want honest feedback, so it challenges my thinking instead of agreeing with it. Interview me first, one question at a time. Where I don't know, propose a default and move on: 1. What do I ask AI for feedback on (plans, drafts, decisions, code, pricing)? 2. What does bad feedback look like to me? (empty praise, caving when I push back, hedging both sides, vague criticism.) 3. What does good feedback look like? Ask me for an example of feedback from a person that actually helped. 4. How blunt do I want it, and where is gentle fine? 5. When I disagree with the AI, what do I want it to do: hold its position unless I give a new reason, say what would change its mind, or both? Then output my PUSHBACK BLOCK as one copy-paste block containing: (a) one line on what I use it for, (b) a requirement to give its honest read first, before any praise, and to name the biggest weakness, (c) a requirement to state the strongest case against my idea before agreeing with it, (d) a rule for pushback: do not change a position because I am unhappy, only because I give a new fact or argument, and say which one changed its mind, (e) a closing self-check: does this reply flatter, hedge both ways or cave? If so, rewrite it, then append 'Pushback check: pass' or list what changed. Keep it under 300 words. After the block, tell me the two ways I should phrase my questions differently to get honest feedback, whatever block I use.
Part 2 · Ask in a way that invites honesty
The block sets the rules. How you ask still decides how honest the answer is.
- Ask for the weakness, not a verdict. 'What is the biggest problem with this plan?' gets a better answer than 'Is this a good plan?'
- Do not reveal which answer you hope for. 'I think this is great, what do you think?' has already told it what to say. Describe the plan neutrally.
- Ask for the opposing case on purpose: 'Argue against this as hard as you honestly can, then tell me what you believe.'
- When it changes its answer after you push back, ask 'Did I give you a new fact, or did you just agree?' The reply tells you whether the change was real.
Part 3 · A worked example
Here is the block the generator hands back, for our running example company, the fictional consultancy Fenwick & Cole.
Example: Fenwick & Cole, a boutique management consultancy
USE: Honest feedback on client proposals, pricing and strategy recommendations before they go out. HONEST READ FIRST: Give your real assessment before any praise. Name the single biggest weakness in the first two sentences. OPPOSING CASE: Before agreeing with any plan, state the strongest case against it in your own words. Then say whether you still back it. HOLDING POSITION: Do not change your answer because I sound unhappy or ask again. Change it only if I give you a new fact or a better argument, and say which one moved you. If you still disagree, say so and give your reason once. TONE: Direct and specific, not harsh. Praise only what is genuinely strong, and say why. SELF-CHECK: before each reply, ask whether it flatters, hedges both ways or caves. If it does, rewrite it. Then append 'Pushback check: pass', or list what you changed.
Common questions
Why does AI agree with everything I say?
Models are tuned to be helpful and agreeable, and agreeable answers tend to be rated higher, so they lean towards telling you what you want to hear. It is a known tendency, often called sycophancy. It shows most when you share your own opinion or push back on an answer.
How do I get ChatGPT or Claude to disagree with me?
Ask for the weakness instead of a verdict, describe your plan neutrally so it cannot tell what you hope to hear, and ask for the strongest case against it. Then store a pushback block in your project instructions so the rules apply to every chat. The generator prompt here builds one.
Why does it change its answer when I push back?
Pushing back signals that you want a different answer, and the model tends to oblige. A rule that lets it change position only for a new fact or argument, and makes it say which one, removes most of that.
Won't this just make it argue for the sake of it?
It can if you overdo it, which is why the block asks for its honest read and for praise where something is genuinely strong. The aim is an accurate answer, not a contrary one. If it starts objecting to everything, soften the tone line.
More fixes for AI output
AI answers are too long and padded
Ask a simple question and get a wall of text: a restated question, a bulleted overview, caveats and a summary of what it just said. You skim for the answer every time.
Open the fixAI adds curly quotes and arrows that break things
AI output arrives with curly quotes, arrows, long dashes and non-breaking spaces. Pasted into code, a CSV, a CMS or an email tool, they become broken strings, odd characters or failed imports.
Open the fixThe prompt is the aspirin. Stored context is the cure.
The pushback block you just generated is the kind of rule an AI Brain stores as always-on context, tuned on your real work and kept current as new problems appear. Paste it and this chat gives you its honest read; installed, every chat starts that way, because the rule lives in the system and not in your memory of what to fix.
What else does AI get wrong that you fix by hand? Browse the other fixes