Former researcher leaks: 10 prompts used internally by Anthropic

God Translation Bureau is a compilation team under 36Kr, focusing on fields such as technology, business, workplace, and life, and mainly introducing new technologies, new ideas, and new trends from abroad.

Editor's note: Stop using Claude as just a chat tool. That's a waste. These 10 prompt frameworks leaked from Anthropic are the real keys to unlocking deep reasoning. This article is a compilation.

Most people only use Claude as a basic chatbot, wasting 60% to 70% of its reasoning ability.

These 10 prompts are the real way experts use Claude. These prompts have all been internally tested and are designed to achieve the highest level of clarity, honesty, and depth. They are ready to use without any nonsense.

Bookmark this article. Your way of using Claude is about to change dramatically. (Professional advice: Use them in order for better results.)

Prompt 1: Context Briefing (The "Map" Claude Really Needs)

Never just throw a question at it. First, provide rich background information:

"You are assisting me in achieving [specific goal]. My background is: [your role + company/project + constraints]. I have already tried [X and Y]. I'm currently stuck at step [Z]. Please confirm that you fully understand the above background before making any suggestions."

Internal tests show that this single change can improve the output quality by 41%. Claude doesn't have mind-reading abilities. Give it complete guidance.

Prompt 2: Force the Display of the Reasoning Process (Enhanced Chain of Thought)

Don't just ask for the answer directly. Ask it to show the process: "Before giving any final suggestions: - Show your complete step-by-step reasoning process. - Clearly list every assumption. - Mark uncertainties and confidence levels (low/medium/high). - Only after completing the above steps, give a refined answer."

This can uncover Claude's hidden reasoning layers. You'll get not only an answer but also an auditable and truly trustworthy thinking process.

Prompt 3: Force Honesty (Reject the "People-Pleasing" Personality)

Claude's training goal is to be as helpful as possible... but sometimes that means it might be tactful to please. Force it to tell the truth: Please be absolutely honest, even if it sounds uncomfortable. If my idea has fatal flaws, please speak bluntly. Don't sugarcoat the language or add polite disclaimers. I want to hear the unpalatable truth now to avoid future failures. This can activate Claude's honesty layer within the Constitutional AI framework. For strategies and planning, this is a game-changing technique.

Prompt 4: Ultra-Specific Role Setting (Abandon the Mediocre "Act as an Expert")

Saying "act as an expert" is too amateurish. Try this: "You are a [very specific role] with [exact number of years] of experience, who has witnessed [specific failure patterns in this field]. Please think using [specified framework or methodology]. Be direct in your speech, skip general suggestions, and only refer to real-world patterns you've 'observed'."

Specificity = surgical precision in every response.

Prompt 5: Devil's Advocate Mode / Red Team Evaluation Mode

Claude has a default tendency to agree with the user. Break this: "I'm about to share a plan/idea. Your only task is to destroy it. Mercilessly point out every wrong assumption, overlooked risk, second-order effect, and possible failure point. Don't hold back. Be my red team." This is exactly how Anthropic internally stress-tests ideas. One prompt can turn Claude from a "cheerleader" into a top-notch critic.

Prompt 6: Lock in the Scope (Eliminate Hallucinations at the Source)

Claude likes to think divergently. Immediately constrain it: "Please answer strictly within [exact scope/background]. If there is content outside this scope, say 'out of scope' and stop answering. I'd rather know what you don't know than confident speculations." Just this one line can significantly reduce those random and baseless statements.

Prompt 7: Lock in the Output Format (Precision Engineering)

Claude defaults to outputting long, flowery prose. Force it to output in the format you need: "Please strictly organize your entire response according to the following structure, without including other content: 1. One-sentence summary. 2. List of key points (maximum 3 - 5 points). 3. A clear next action. Please use Markdown format. No extra words." Claude follows format instructions more faithfully than any other model. Make good use of this.

Prompt 8: Assumption Audit (Reality Check After the Answer)

After getting any complex answer, immediately run this: "List all the assumptions you made in this response that I need to verify in the real world. For each assumption, explain: (a) what would happen if the assumption is wrong, and (b) how the suggestion would change accordingly." Most plan failures stem from implicit assumptions. This prompt can expose them.

Prompt 9: Compression Cycle (A Godsend for Long Context Processing)

In a long conversation, execute this every 5 - 6 messages: "Please strictly summarize our current progress in the following format: • Problems solved. • Decisions made. • Most important pending problems. • Suggested next focus points." This can prevent Claude from going off-topic and avoid an overly heavy context burden, keeping your long conversations highly focused.

Prompt 10: Pre-Mortem (Anthropic's Secret Weapon)

Before starting any project: "Assume that this project/idea fails in 6 months. Please write a detailed 'post-mortem' as if the failure has already happened. List the top three most important reasons for the failure in order of probability. Please be specific and harsh." Anthropic's product team runs this process for every major decision. It can catch blind spots that other review methods can't.

Translator: boxi.