"Steal This Prompt (Fix)"...

Most people write: context → backstory → finally the actual ask. Flip it. Lead with the exact output you want, then give context. His tests showed a 31% improvement in response accuracy on complex tasks just from reordering.

"Could you please help me with..." costs you reasoning tokens. Claude allocates its attention budget before generating a single word. Filler phrases eat into that budget. Every unnecessary word is a reasoning token you're not getting back.

"You are an expert marketer" = weak signal. "You are an expert marketer. Never use adjectives that can't be measured." = strong signal. The constraint forces Claude to activate the role instead of just wearing it as a costume.

Telling Claude what you don't want works better than telling it what you do want. "Don't write in a formal tone" outperforms "write casually" in 7 out of 10 tests. Negative framing activates a completely different reasoning pathway.

Don't describe the format. Start writing it. Instead of "give me a numbered list," write: "Here are the key findings: 1." Claude will follow the pattern. Every time. No exceptions.

Add this line to any factual request: "Flag anything you're less than 90% confident about." Claude's hallucination rate on complex topics drops by over 40% with this single addition. Most people never use it.

Don't ask Claude to do 4 things in one message. Ask it to complete step 1, confirm, then proceed. Each checkpoint resets Claude's attention and prevents context drift on long tasks.

Before any analysis task, add: "First, make the strongest possible case for the opposite view." This forces Claude out of confirmation mode and into actual reasoning. The quality difference is not subtle.

"Before responding, silently identify the 3 most likely ways you could misinterpret this request." Claude never shows you this check. But it runs it. Response accuracy on ambiguous prompts jumps significantly.