Most internet advice on writing Claude prompts is either toy-example demos or 14-step "prompt engineering" frameworks for use cases you do not have. This is the real-world version — patterns we use when building Claude Projects for B2B teams, with specifics on what to include and what to leave out.
The single most effective technique: load 5-10 strong examples into the Project's knowledge, and have the system prompt say "Match the voice and structure of the examples in your knowledge." This works better than any amount of adjective-based voice description.
If you must describe voice in words, be concrete and behavioral: "Use short declarative sentences. Avoid adjectives that praise the reader. Use second person." Beats "Be confident and warm."
Voice is shown, not told. Two paragraphs of example output do more than two pages of voice guidelines.
Be explicit about format. Specify:
Almost every production prompt benefits from explicit forbiddances. Common ones:
You are a [role] at [Company], a [one-sentence company description]. Your job is to [specific task]. Your audience is [specific persona], who care about [their priorities]. Match the voice and structure of the example [outputs] in your knowledge — direct, evidence-forward, [adjective specific to your brand]. Output as [format], with sections [list]. Each section should be [length]. Do not [forbidden behavior 1]. Do not [forbidden behavior 2]. If you are missing [critical input], ask for it before producing output.
Replace every bracketed phrase with your specifics. Test with 3-5 real inputs. Refine the prompt based on what fails. Save the working version in your Project. That's it.