How to prompt Claude Fable-5 (and any AI model): → Start by deleting the long examples you give. → Delete the "step-by-step" prompt technique. → Avoid biased questions. Don't finish with "right?" → Chain-of-Thought (step-by-step) only works well with maths problems. It's worth skipping it. → Don't give a long list of rules (max 3) to follow. The models got very good at simple prompts, and worse at your complicated, over-structured ones. To make a simple prompt, better: → Make your next prompt shorter. → Define the goals with extreme clarity. → Give no more than 3 rules to perform one task. → Ask open-ended questions to let it figure it out. → Always encourage it to ask clarifying questions. → Don't trust the sources blindly. It tries to prove itself by finding links that agree with what it already has written. This is the one simple prompt template I stick to: "[GOAL: What a finished answer must achieve. Write the test it has to pass.] [CONTEXT: The facts the AI would otherwise invent. Only the ones it cannot know.] [CONSTRAINTS: The hard limits. 3 max.] If anything here is missing or unclear, ask me before you answer. [Add “use AskUserQuestion tool” if you’re on Claude]." I've been teaching prompt engineering since 2022, and so much has changed since then. What was true - because science said so - isn't anymore. The models are 10x more capable now. I'm a psycho so I actually filter out and read the 100+ daily papers on prompt engineering (on arxiv. org). Read my findings at