Anthropic prompting guide distilled

Anthropic’s prompting advice has been circulating again through a third-party distillation that turns the company’s broader guidance into a short checklist for Claude users. The writeup centers on a simple claim: better prompts often matter more than switching models when users want more reliable output. Anthropic’s own documentation supports much of that framing, though it does not present the advice as the same nine-rule list. The company’s public materials instead spread the guidance across best-practices pages, notebooks and an interactive tutorial. (platform.claude.com) The practical overlap is clear. (github.com) Anthropic says its current prompting documentation covers output control, thinking, tool use and agentic systems, and its GitHub tutorial is organized into nine chapters with exercises plus an appendix. Those materials teach users to be clear and direct, assign roles, separate data from instructions, format output explicitly, ask for step-by-step thinking, use examples and reduce hallucinations. ### Where did the “31 pages” claim come from? A May 10 post on Dev.to said the author had read “31 pages of Anthropic prompting guidance” and pulled out the changes that “actually” matter for Claude Opus 4.7. (dev.to) The post is not an Anthropic publication, but it closely tracks themes Anthropic highlights in its own documentation about clarity, formatting, verbosity control and prompting for thinking. Anthropic’s official best-practices page describes itself as the “single reference” for prompt engineering with Claude’s latest models. That page says it covers foundational techniques, output control, tool use, thinking and agentic systems for models including Claude Opus 4.7. (platform.claude.com) ### Why does the checklist focus so heavily on output format? Anthropic’s tutorial devotes one full chapter to “Formatting Output and Speaking for Claude,” and the Dev.to distillation pushes the same point in blunt terms: name the table, columns, bullets or structure you want, or Claude will choose one for you. In the example cited there, a vague request to review a contract becomes a structured table with clause, risk, severity and rewrite fields. (dev.to) Anthropic’s best-practices page also says users may need to tune prompts if their product depends on a certain style or verbosity of output. The company says Claude Opus 4.7 calibrates response length to how complex it judges a task to be, rather than defaulting to a fixed level of detail. (platform.claude.com) ### What is behind the advice to specify length and tone? Anthropic says response length is no longer something users should leave implicit. Its docs recommend adding direct instructions such as asking for concise, focused responses and skipping non-essential context when shorter output matters. (dev.to) The Dev.to summary makes the same point more operationally: if the input is long, the summary may also run long unless the prompt sets a cap such as “exactly 5 bullets” or a word limit per bullet. The post also says Claude 4.7 has a “colder default tone” than earlier versions and that users who want warmth should ask for it directly or provide reference sentences. (platform.claude.com) ### Why do positive instructions matter more than “don’t” commands? Anthropic’s docs say positive examples showing the desired level of concision tend to work better than negative examples or instructions telling the model what not to do. (platform.claude.com) That is one of the clearest official matches to the third-party rule to replace “don’t use jargon” with concrete directions about what language to use instead. The Dev.to post goes further, arguing that every negative instruction should be rewritten as a positive instruction plus a concrete swap example. Anthropic’s tutorial, while older, is built around the same broader idea that specificity and examples steer output more reliably than vague constraints. (dev.to) ### Why does step-by-step prompting keep showing up? Anthropic’s public tutorial includes a chapter called “Precognition: Thinking Step by Step,” making explicit reasoning one of the core skills it teaches. The company’s current best-practices page also includes “thinking” among the central areas of prompt design for Claude. (platform.claude.com) A separate Ruben Hassid cheat sheet surfaced similar language, including a recommendation to ask for chain-of-thought-style work in multiple steps and to tell the model to work step by step. That document is not an Anthropic source, but it shows how the company’s guidance is being repackaged for mainstream users. (dev.to) ### Where can readers check the source material? Anthropic’s public best-practices page and its GitHub tutorial remain the clearest primary sources for the company’s prompting guidance. The GitHub repository says the course is broken into nine chapters with exercises and an appendix, and Anthropic’s docs page links users to model-specific updates, migration guidance and prompt-engineering patterns for current Claude releases. (github.com 1) (github.com 2) (assets.super.so)