(Why these skills exist)
Agent-generated design fails
in four predictable ways.
You think your agent's design problem is prompt quality. It isn't. The agent has no way to judge its own output, so it ships the first thing that works. Every era of design solved this with a named standard. Ours had none. CRISP is that standard.
I · 1446 BC
It doesn't know who it's building for.
"I have filled him with wisdom, with understanding, and with knowledge, and with all kinds of skill."
Exodus 31:3, on Bezalel, the first named designer
The oldest design standard on record is not about skill. Bezalel's third gift, da'at, means knowing who the work is for and why. An agent dropped into your codebase has none of it. So it builds for nobody in particular. Empty states say "No data available". Headings could belong to any product. One settings page serves the first-week beginner and the five-year expert alike.
The fix: give the agent that knowledge once. /crisp-teach asks you about your users, your design system and what good looks like for your product. It saves the answers in .crisp.md, so every later command starts with that context. /crisp-brief turns a vague request into a clear brief with success criteria you can check. /crisp-research shows what competitors do before you start designing.
II · 25 BC
Functional is not the same as good.
"Well building hath three conditions: firmness, commodity, and delight."
Vitruvius, De Architectura
Code that compiles meets only the first condition, firmness (firmitas). Agents stop there. The interface works. But it shows a spinner every time you change a filter. It shows "1,247" with no comparison and no meaning. It sends you to another portal to approve one item. Technically correct. Hollow to use.
The fix: a real evaluation. /crisp-audit scores all five dimensions, rates each problem from P0 (critical) to P3 (minor) and lists the fixes in priority order. /crisp-review gives a grade and the top three issues in thirty seconds. /crisp-ux-laws and /crisp-a11y check the work against proven rules of how people think and the WCAG 2.2 AA accessibility standard.
III · 1976
Agents add. They never subtract.
"Good design is as little design as possible. Less, but better."
Dieter Rams
Left alone, an agent answers every request by adding something: another toggle, another card, another pop-up. Soon all twenty-three settings show to every user. Rams built his ten principles on the opposite idea: good design is about what you take away.
The fix: make removal a command. /crisp-redesign updates an existing UI without breaking what already works. /crisp-copy cuts interface text down to the words that help people find their way. /crisp-design-eng checks motion and polish, so nothing moves without a reason.
IV · 2026
Nothing survives the session.
"Your AI agent can generate code. It can't evaluate design."
CRISP
Close the session and the context is gone. The next session decides everything again: a new spacing scale, a different empty state, a fourth shade of grey. For three thousand years, people wrote design standards down so the knowledge would outlive the designer. Your agent's knowledge is gone in hours.
The fix: a written record the agent keeps up to date. .crisp.md keeps your context across sessions and updates. /handoff turns a reviewed design into a spec another agent can build from. /feature-design records why each flow decision was made. /crisp-ai applies the standard to AI features such as chat and agents, and /crisp-funnel builds a full funnel to the same standard.
Bezalel · Vitruvius · Rams · CRISP. 3,472 years of one question: does this design actually work for people?