🚨BREAKING: A former Anthropic researcher just leaked the internal Claude prompting playbook.
No speculation. No guesswork. Direct from someone who built it.
You're burning 35% of Claude's reasoning capacity with one common mistake.
Here are the 10 prompts they use that nobody outside Anthropic knows about:
Prompt 1: The Situation Brief
Never start with the ask.
Start with the situation:
"Here's my context: [role, company, problem]. Here's what I've already tried: [X, Y]. Here's where I'm stuck: [Z]. Now help me think through this."
Internal testing showed this single change increases useful output by 41%.
Claude needs the map before it can navigate.
Never start with the ask.
Start with the situation:
"Here's my context: [role, company, problem]. Here's what I've already tried: [X, Y]. Here's where I'm stuck: [Z]. Now help me think through this."
Internal testing showed this single change increases useful output by 41%.
Claude needs the map before it can navigate.
Prompt 2: The Reasoning Demand
Don't ask for the answer. Ask for the thinking.
"Before you give me a solution, walk me through your reasoning step by step. Show me where you're uncertain. Flag any assumptions you're making."
This forces Claude's internal reasoning layer to surface.
What you get back isn't just an answer. It's a thought process you can interrogate.
Don't ask for the answer. Ask for the thinking.
"Before you give me a solution, walk me through your reasoning step by step. Show me where you're uncertain. Flag any assumptions you're making."
This forces Claude's internal reasoning layer to surface.
What you get back isn't just an answer. It's a thought process you can interrogate.
Prompt 3: The Honest Constraint
Claude defaults to helpful.
Sometimes helpful means telling you what you want to hear.
Override it:
"I need you to be honest even if it's uncomfortable. If my plan has a fatal flaw, say so directly. Don't soften it. I'd rather hear the hard truth now than fail later."
This unlocks Claude's constitutional backbone.
Claude defaults to helpful.
Sometimes helpful means telling you what you want to hear.
Override it:
"I need you to be honest even if it's uncomfortable. If my plan has a fatal flaw, say so directly. Don't soften it. I'd rather hear the hard truth now than fail later."
This unlocks Claude's constitutional backbone.
Prompt 4: The Role Specification
"Act as an expert" is the weakest prompt in existence.
Use this instead:
"You are a [specific role] with [specific experience] who has seen [specific failure mode] before. You think in [specific framework]. You are direct and skip conventional advice."
The more specific the identity, the more specific the reasoning.
Vague persona = vague output. Every time.
"Act as an expert" is the weakest prompt in existence.
Use this instead:
"You are a [specific role] with [specific experience] who has seen [specific failure mode] before. You think in [specific framework]. You are direct and skip conventional advice."
The more specific the identity, the more specific the reasoning.
Vague persona = vague output. Every time.
Prompt 5: The Devil's Advocate
Claude is trained to be agreeable.
Break that pattern:
"I'm about to share my plan. Your job is to destroy it. Find every assumption that could be wrong, every risk I'm ignoring, every reason this fails. Don't hold back."
This is how Anthropic's own teams pressure-test ideas internally.
A model that challenges you is worth 10x one that agrees with you.
Claude is trained to be agreeable.
Break that pattern:
"I'm about to share my plan. Your job is to destroy it. Find every assumption that could be wrong, every risk I'm ignoring, every reason this fails. Don't hold back."
This is how Anthropic's own teams pressure-test ideas internally.
A model that challenges you is worth 10x one that agrees with you.
Prompt 6: The Scope Pin
Claude will answer the question you asked AND three adjacent questions you didn't.
Pin the scope early:
"Stay strictly within [X context]. If something falls outside this scope, tell me rather than speculating. I'd rather have a gap than a confident wrong answer."
This kills hallucination at the source.
Claude stops filling gaps with plausible fiction.
Claude will answer the question you asked AND three adjacent questions you didn't.
Pin the scope early:
"Stay strictly within [X context]. If something falls outside this scope, tell me rather than speculating. I'd rather have a gap than a confident wrong answer."
This kills hallucination at the source.
Claude stops filling gaps with plausible fiction.
Prompt 7: The Format Command
Claude writes for comprehensiveness by default.
You probably need brevity.
Set the format upfront:
"Structure your response as: 1) One sentence summary. 2) Three key points. 3) One recommended next action. Nothing else unless I ask."
Claude follows format instructions more precisely than any other model.
Use that precision deliberately.
Claude writes for comprehensiveness by default.
You probably need brevity.
Set the format upfront:
"Structure your response as: 1) One sentence summary. 2) Three key points. 3) One recommended next action. Nothing else unless I ask."
Claude follows format instructions more precisely than any other model.
Use that precision deliberately.
Prompt 8: The Assumption Audit
After any complex answer, run this:
"What assumptions did you just make that I should verify? What would change your answer if those assumptions were wrong?"
The most underused prompt in existence.
It surfaces the hidden scaffolding under every response.
The scaffolding is where most plans fall apart.
After any complex answer, run this:
"What assumptions did you just make that I should verify? What would change your answer if those assumptions were wrong?"
The most underused prompt in existence.
It surfaces the hidden scaffolding under every response.
The scaffolding is where most plans fall apart.
Prompt 9: The Compression Loop
For long sessions, Claude accumulates context debt.
Every 5-6 exchanges:
"Summarize where we are. What problem are we solving, what have we decided, and what's the single most important thing we haven't resolved yet?"
Keeps the session locked on what matters.
Prevents Claude from solving the wrong problem confidently.
For long sessions, Claude accumulates context debt.
Every 5-6 exchanges:
"Summarize where we are. What problem are we solving, what have we decided, and what's the single most important thing we haven't resolved yet?"
Keeps the session locked on what matters.
Prevents Claude from solving the wrong problem confidently.
Prompt 10: The Pre-Mortem
Before you ship anything Claude helped build:
"Assume this fails 6 months from now. Walk me through the 3 most likely reasons why. Be specific. What would the failure actually look like?"
This is the prompt Anthropic product teams run on every major decision.
It catches what every other review process misses.
Before you ship anything Claude helped build:
"Assume this fails 6 months from now. Walk me through the 3 most likely reasons why. Be specific. What would the failure actually look like?"
This is the prompt Anthropic product teams run on every major decision.
It catches what every other review process misses.
I hope you've found this thread helpful.
Don't forget to bookmark for later.
Follow me @heyrimsha for more.
If you enjoyed reading this post, please support it with like/repost of the post below 👇
Don't forget to bookmark for later.
Follow me @heyrimsha for more.
If you enjoyed reading this post, please support it with like/repost of the post below 👇
View Tweet
Generated by Thread Navigator
Press ⌘ + S to quick-export