Make the Other Mistake (Prompting for Brutal Feedback)
To get honest AI critique, over-correct toward brutal — the model won't actually overshoot.
- Difficulty
- Starter
- Time to result
- ~days to results
- Steps
- 4
- Confidence
- 90%
Language models default to being agreeable and rarely become as harsh as a genuinely blunt human. To extract useful criticism, deliberately over-steer the instruction toward the opposite of your own tendency. If you're naturally too nice, ask the model to be brutal — it will still land only moderately critical, which is where you actually want it.
Origin
Mike Krieger credits 'a great essay' on the idea of deliberately 'making the other mistake' when you know you have a directional bias. He applies it to prompting Claude as a product-strategy thought partner.
Core principles
- 01Models are trained to be pleasant and lean toward agreeableness
- 02An instruction to be 'brutal' still resolves to merely 'more critical' in practice
- 03Correcting toward your opposite bias lands you near the true center
- 04Vague asks ('what could be better?') get vague, hedged answers
How to run it
- 1
Name your own bias
Recognise the direction you personally lean — usually too polite, too optimistic, or too attached to your own idea.
Pro tip The same logic works in reverse: if you tend to be harsh, ask for a balanced or generous read.
- 2
Over-steer the instruction to the opposite extreme
Instead of asking 'what could be better on this?', instruct the model to 'be brutal, roast this, tell me what's wrong with it.'
Pro tip Frame it as a role: 'What thought patterns have I fallen into that you want to break me out of?'
Watch out The model still won't become the most critical person alive — you are compensating for its niceness, not creating a monster.
- 3
Add a reasoning nudge when depth matters
For harder critiques, prompt it to 'think hard' so it engages a more deliberate reasoning flow before answering.
Pro tip In Claude Code, 'think hard' is literally reacted to and changes the reasoning budget.
- 4
Read the output as calibrated, not literal
Treat the resulting critique as landing near the honest center rather than at the harsh extreme you requested, and mine it for the genuinely new angles.
In the wild
Krieger previously asked Claude soft questions like 'what could be better on this product strategy?' and got polite editing. He switched to 'just roast this product strategy' and 'be brutal Claude, tell me what's wrong with this strategy.'
→ The harsher framing forced Claude to be meaningfully more critical, making it a more useful strategy reviewer even though it still couldn't be pushed to be truly savage.
Common mistakes
Asking for feedback in soft, open-ended terms
Prompts like 'what could be improved?' let an agreeable model hedge and reassure, so you never surface the real weaknesses.
Expecting the model to overshoot into cruelty
Because the model won't actually become maximally harsh, a mild 'be a bit more critical' nudge under-corrects — you must ask for the extreme just to reach the middle.
Is it for you?
Best for
Anyone using an LLM as a thought partner on strategy, writing, or decisions who keeps getting flattering, unhelpful validation.
Not ideal for
Sensitive personal contexts or situations where you need supportive framing rather than adversarial critique.
From the transcript
“there's a great essay around like make the other mistake like if you tend to be too nice can you focus on like even if…”
“um so with Claude sometimes I'm like be brutal Claude roast me like tell me what's wrong with this strategy”
“like what are the thought patterns that I've like fallen into that you want to break me out of”
From the episode
Anthropic’s CPO on what comes next
Mike Krieger (co-founder of Instagram)