/crisp-ai
A standard for AI surfaces.
What it does
A design standard for AI surfaces. /crisp-ai evaluates chat interfaces, streaming responses, AI agents, generative UI, and inline assist against six AI-specific dimensions — problems that did not exist before 2023, from communicating uncertainty without meaningless percentages to keeping streaming text readable.
Each dimension maps back to the core CRISP framework, scores out of ten, and produces named violations with exact fixes — plus component-level recommendations benchmarked against products like Claude.ai, Perplexity, and Cursor.
When to reach for it
Reach for it when AI is a primary interaction, not a garnish:
- Building or reviewing a chat interface, agent surface, or generative UI
- Adding an AI feature to an existing product and deciding how it should behave when it is wrong
- Streaming output feels disorienting and you need to name why
- Users either over-trust or ignore the AI and the interface offers no calibration
What it produces
- 6-dimension AI UI scorecard
- Streaming UX recommendations
- Agent feedback pattern guidance
- Error + uncertainty handling
In the terminal
Illustrative example output — not a real audit.
The six dimensions
- Uncertainty communication — does the user know when to trust and when to verify?
- Streaming legibility — is partial output readable as it arrives?
- Failure grace — when the AI fails, is recovery clear without blaming the user?
- Human override clarity — can the user correct, edit, or undo AI output easily?
- Context transparency — does the user know what context the AI is using?
- Progressive AI power disclosure — are capabilities introduced without overwhelming?
Common questions
How does this relate to /crisp-audit?
They are complementary scorecards. /crisp-audit scores the five core dimensions; /crisp-ai scores the six AI-specific ones. An AI-native product needs both — and if .crisp.md marks the product AI-Native, this evaluation runs as part of the full audit automatically.
What surfaces does it cover?
Chat interfaces, inline assist inside editors, autonomous agents, generative UI, AI-powered search, and hybrid surfaces where AI is embedded in a conventional product. It asks which type you are working on first, because the failure modes differ.
Why not just show confidence percentages?
Because users do not calibrate them. The uncertainty dimension pushes qualitative patterns instead — source citations, verify prompts, visual de-emphasis of generated content, and surfacing knowledge cutoffs when they matter.
Does it assume the AI will fail?
Yes, deliberately. AI fails in distinctive ways — context overflow, content policy, knowledge limits, confident wrongness — and each needs its own recovery path. Designing around failure honestly is the standard, not hiding it.
Install
Installs all fourteen CRISP skills and auto-detects your AI harness. Run /crisp-teach once per project first.
Related skills