A 20-minute structure
- Minute 0-3: name the feature surface, source, and safety boundary.
- Minute 3-8: run the prompt and narrate the setup rather than apologizing for it.
- Minute 8-13: inspect the result with the verification checklist.
- Minute 13-17: show one variation or backup prompt if the anchor prompt needs a retry.
- Minute 17-20: list next experiments, risks, and what the demo did not prove.
Choose the anchor demo
For a broad team, a clickable artifact prototype works because the result is visible and can use synthetic data. For an analytical room, a PDF evidence table or web research brief is better because the verification step matters as much as the reveal. For a design room, a screenshot audit turns visual evidence into a practical critique.
Name the model out loud
Say which model you are running before the reveal, because the first question after a good demo is "can I do that on my account?" As of September 2026 the generally available lineup is Claude Fable 5.1, Claude Opus 5, Claude Sonnet 5, and Claude Haiku 4.5. Most demos in this gallery are cast on Opus 5, which is the default on Claude Max and the strongest model on Pro, so the room can usually reproduce the run without API access. Seethe casting page for the exceptions.
Naming the model also protects you later. A demo recorded without a model name becomes unreproducible the moment the lineup changes, and it changed twice in the last three months.
Use source-backed language
Say "Anthropic documents artifacts as a way to create shareable apps, tools, and content" rather than "Claude can build anything." Say "web search responses include citations" rather than "Claude always knows what is current." The first version can be checked; the second invites mistrust.
End with the anti-claim
The anti-claim is the sentence that preserves credibility: this demo does not prove production readiness, full accuracy, or policy approval. It proves a pattern is worth inspecting more carefully.