A good Claude demo prompt shows the strings

The reusable structure

Anthropic's prompting guidance emphasizes explicit instructions, output formats, tool-use direction, and model-specific behavior. MagicClaude turns that into a stage-ready pattern: say what Claude is, what it should reveal, what evidence it can use, and how the viewer should verify the result.

<role>Give Claude a job title that matches the reveal.</role>
<goal>State the artifact, analysis, or workflow the audience should see.</goal>
<context>Add only the data Claude needs. Keep private data out of public demos.</context>
<requirements>
- Ask for assumptions before final output.
- Define the output shape.
- Require uncertainty labels or verification steps.
- Ask for a short "what this does not prove" note.
</requirements>
<source_boundary>Use only uploaded material and cited web sources when factual freshness matters.</source_boundary>

Four moves that make the reveal credible

1. Name the job. "Act as an evidence clerk" gets different behavior than "summarize this."
2. Ask for assumptions first. This keeps the first answer from looking more certain than it is.
3. Define the output shape. Tables, checklists, and schemas are easier to verify than prose alone.
4. Include the anti-claim. "What this does not prove" is the simplest way to avoid demo inflation.

When to require sources

If the claim depends on current facts, ask Claude to use web search or Research and cite sources. If the claim depends on an uploaded document, ask for page, section, or quotation anchors and verify them in the original.