The Skill Prism

Resolving one prompt into a spectrum of skills and subagents

Have you ever popped open ChatGPT or Claude ("agent" from here on) and asked it to draft a document? A legal first draft, a proposal for a client, or some code? There's a gap between "let's draft a document" and an output you'd actually use. Closing that gap requires a perspective shift. I view that shift through an analogy of light. A bare-bones agent is like white light, every color at once, undifferentiated, nothing resolved. Apply a prism to it, though, and you can scatter that light into distinct wavelengths, skills, to reliably produce the output you’re after. Let's scatter that light.

Imagine an empty agent prompt, a blinking cursor, a conversation that can go anywhere. It can be a chat about movies, EU legislation, or ideal plants for your gardening zone. Everything exists in that blinking cursor. That is the white light. When you’re doing knowledge work, this white light approach breaks down. You need output with actual confidence behind it. That confidence comes from applying the prism to that blinking cursor, repeatedly, as you learn to see the muddiness in the output.

Now a concrete example: writing code. There is a tremendous breadth of experience underlying this. It’s not just writing code. It’s gathering context in meetings. It’s understanding “agile” practices. It’s evaluating a path through the code when starting fresh work. It’s beginning with Test-driven Development. Or not. It’s planning. Then it’s writing. Then it’s reviewing. Asking an agent to “write code” glosses over a tremendous amount of experience to arrive at a final, valuable output. No matter the profession, the prism analogy teases out these processes into more granular refractions.

Testing, reviewing, writing are the refractions we’re chasing. Imagine them as individual, curated skills that capture your experience. A skill defining what a good test looks like. What a bad test looks like. A skill outlining what a good agile story looks like and how to compose one. A skill to trace change surfaces of a feature request right in front of you, the way you do yourself; it informs your next step, building context the way you do. They focus on one thing in order to do it well. They’re the spectrum of your experience.

The prism doesn’t just work for skills. That’s where it starts. Once you’re fluent in seeing the colors, the same prism splits your agent into subagents, each representing one color, owning its domain.