Build Your AI Use Like a Design System

You use AI tools every day, but every session feels like starting from scratch. The prompt that worked yesterday, you have to rewrite today because you forgot it. Third time doing a similar task and you're still figuring out the sequence from zero. This essay is for you. We'll go slowly.

Up front: structure your AI use like a design system. Making it fresh every time vs. building a system once produces a 10× difference in three months. Today I'll show you the three floors.

Quick concept. Anyone who's used Figma has heard of "design systems." Design teams build them to ship consistent output fast. Three floors.

Floor 1. Tokens — smallest unit. One color, one font size, one spacing value. Named atoms, like "Primary Blue = #0066CC."

Floor 2. Components — reusable pieces built from tokens. Buttons, cards, input fields. Build once, call anywhere.

Floor 3. Patterns — situational uses that combine components. "Login pages look like this," "Dashboards have this layout."

With these three floors, design is consistent, fast, and easy to change. Exactly this structure can be applied to AI use. That's the core of today.

Floor 1. AI tokens are reusable prompt fragments. Not the whole prompt, just a fragment.

I use four token categories.

Role tokens: "You're a 10-year editor," "You're a writer for marketers in their 40s." Tells AI the stance to answer from.

Context tokens: "This project is B2B SaaS onboarding," "The reader is a startup founder." Project background fragments.

Format tokens: "Return as 5 bullet points," "Output as a markdown table." Format fragments.

Tone tokens: "Friendly first-person," "Academic third-person," "No casual register." Voice fragments.

Make 5-10 of each category and store them in a note. When you need one, combine them. [Editor role] + [B2B context] + [5-bullet format] + [friendly tone] = one prompt. You don't write from zero every time.

Floor 2. AI components are repeated workflows. Combining tokens into a complete procedure.

My example: I have a "meeting notes → action items → calendar" component. Three steps as one bundle.

Step 1: Paste raw notes into Claude, call [Role: PM] + [Format: owner/task/deadline] tokens, extract action items.

Step 2: Filter to just mine.

Step 3: Auto-add to Google Calendar.

That three-step is one component. After a meeting, I call "meeting component" and it's done in 10 minutes. No fresh thinking required.

The criterion for making a component: 3+ times a week. Less than that, not worth componentizing. Pull out 3-5 such tasks, build them into components, and hours appear.

Floor 3. Patterns are situational modes. The same person must use AI differently in different situations.

I run two modes.

Expand mode (Brainstorm): When widening ideas. "Pull 10 possible angles on this topic," "Find 3 connections I missed." Ask AI to widen. Token combinations can be loose here — volume is the point.

Converge mode: When you already have too many ideas and need to pick or trim. "Of these 5, pick the strongest and explain why," "Cut 10% of flab from this draft." Ask AI to narrow. Criteria must be explicit.

Alternating expand and converge is the heart of the pattern. Ideation stage → expand. Editing stage → converge. Same AI, two modes.

Analogy: a chef's kitchen. Walk into a well-run kitchen and you see three layers.

Ingredient drawer (tokens): Salt, pepper, soy, sugar — each labeled, each in its place. You don't have to remember the name. You pull it when needed.

Base sauces (components): "Basic tomato," "teriyaki" — prepped in advance. You don't mash tomatoes from scratch every service.

Menu structure (patterns): "Lunch set looks like this," "Dinner course flows like this." Situational plans.

Amateur cooks raid the fridge every time and build every sauce from zero. Pros set up systems and assemble. AI use is the same. Build the system in three months, and from then on you assemble.

The second aha moment:

Three months in, the system-builder and the scratch-starter are 10× apart.

Because the scratch person repeats the same work newly every time. 80% of time goes into "how do I do this again." Only 20% produces output. The system person has "how" already decided and spends 80% on output.

Concretely: setting up 20 tokens, 5 components, and 2 patterns took me about 2 weeks. For the next 3 months, my work speed went up 3-4×. Two weeks invested for three months at 4× speed — that's the system ROI.

Three practical steps.

Step 1. Copy your last week's 10 prompts and lay them out. Common fragments will appear. Those are your token candidates. Pull 5-10, name them, save to a note.

Step 2. Pick 2-3 tasks you do 3+ times a week. Write the sequence as a document. That's a component. Next time, open the doc and execute.

Step 3. Distinguish "expand / converge" modes. Decide which mode this situation needs. Same AI — completely different prompting by mode.

Run these three steps over two weeks, and three months later you'll have your own AI working system.

Summary.

Don't build AI use fresh each time — structure it like a design system. Tokens (prompt fragments) + Components (workflows) + Patterns (situational modes) = your AI working system. Like a chef prepping the kitchen, organizing AI use in three floors makes work 3-4× faster.

Three words: Token. Component. Pattern. The skeleton of the system.

Invest 30 minutes tonight. Copy last week's 10 prompts. Name 5 shared fragments. Those are your first tokens. Tomorrow, one component. Day after, another. Two weeks in, a small AI kitchen is standing. The daily fridge-raid ends that same day.