Anthropic's guide puts clear wording ahead of a role

Start here

A reader who wants Claude to do one task well, rather than experiment for its own sake, is the audience Anthropic's own guide names first. Its overview page, retrieved 16 September 2026, tells the reader to establish success criteria and a first-draft prompt before opening the techniques page at all, and warns that not every failing result is best solved by prompt engineering: sometimes a different model choice fixes latency or cost more directly than a better prompt does. That boundary is stated by the vendor itself, not inferred from outside.

What the documents say

The detailed reference the overview page points to, titled Prompting best practices, is organised in three parts: guidance specific to a single named current model, general techniques that apply across the current line, and migration notes for prompts written for earlier models. Within the general section the stated order is clarity first, then adding the reasoning behind an instruction, then structured examples, then XML tags, then setting a role in the system prompt. The clarity section states a golden rule: show the prompt to a colleague with minimal context and see where they would be confused. The examples section recommends three to five examples, wrapped in tags, and calls them diverse so the model does not copy an accidental pattern instead of the intended one. The role section comes last among the general techniques and says even a single sentence in the system prompt makes a difference. Both pages are Anthropic's own documentation describing its current product, not an outside study of it.

Check this

The concrete check is to apply the colleague test literally: hand a saved prompt, with no other context, to someone who does not know the task, and see where they hesitate. A second check is structural: open the best-practices page and find its model-specific table before reading the general section, since the page states that a technique naming a specific model should be treated as measured on that model and re-checked before applying it to another.

What holds and what fails

The ordering, clarity and structure before role-play, is a useful default because it separates what the model is asked to do from how it is asked to sound, and the guide treats the second as a smaller lever; this reading is editorial, since the page ranks techniques by presentation order rather than by measured effect size. What fails is assuming today's model-specific notes transfer forward: the guide's own migration section exists because prompts written for earlier Claude models are stated to behave differently on current ones.

• Run the colleague test on one saved prompt before adding a role or persona to it.
• Wrap any examples already in a prompt in explicit tags and check whether output consistency changes.
• Note which current model a prompt was tuned on, since the guide keeps a separate section for moving it to another.

Read next to a second vendor's guide, the overlap is the real signal: when two competitors independently put plain, specific wording ahead of stylistic dressing, that is a stronger claim than either page makes alone.