One system, so the product could keep changing.
Design Systems
Sessions designed to assemble themselves, screens stopped being fixed. The design system: tokens, blocks, and behaviour for a product whose content does not exist until a learner opens it.
- Role
- Head of Product Design
- Context
- AI Agent Era
- Scope
- Design System
- Years
- 2025 — 2026
A system’s job is to accommodate the next iteration.
Most design systems are an agreement about appearance.
That is true until the product can think for itself and does something the system did not anticipate. Instead we need to trust and give the context over to those who create to deliver a supior experience without need to correct every pixel.
The test is not whether the system covers today’s screens. It is whether Content Creators or AI Agents can output desireable interactions.
One token set — Colour, type, spacing and motion — one source of truth
Zero redesigns — New exercise types ship on blocks that already exist
Design System Refresh (2026)
Single Source of Truth.
The earlier system had grown by rapid acceleration through years of qualative prototypes and growth hackinh. With different interpretations of the same button. The refresh pulled colour, type, spacing and motion back to a set of tokens athat is reuseable, keeping only what more than one surface actually used.
- One token set behind colour, type, spacing and motion
- Components reduced to what is genuinely used
- Compontents with one purpose but versitile
Modular Exercise System (2025)
Learning types as blocks the teams and AI can order freely.
Every new exercise used to arrive as a new flow, which meant every new exercise added layer of complexity. Making each learning type a self-contained block: its own entry, practice, feedback and exit. Meant the product could sequence them without asking anyone to redesign the session around them.
- Each exercise type self-contained
- Sessions composed from blocks rather screen
- New exercise types ship without a redesign
Actions (2025)
Sessions the product assembles, and a learner still recognises.
Action is the agentic layer where the product deciding what this learner should do now, from the blocks available. The design work was making an assembled session feel authored: consistent rhythm, predictable exits, and enough visible reasoning that adaptation reads as attention.
- Generated sessions holding a consistent rhythm.
- Visible reasoning, the learner can see why this, now
- Every adaptive surface has a way out and a way back
The system stopped being “make these screens match” and became “make anything we generate still feel like us”. — Design principle, 2025
Designing components for content that does not exist yet.
With generated content there is no real thing to preview, so every component had to answer four questions before it shipped:
- What is the primary learning action for the learner?
- Does another compotent already achieve the intend outcome?
- Can the compotent be used more than once?
- How else might this compotent be used?
A system that absorbs change instead of resisting it.
New exercise types ship without a redesign, and AI surfaces inherit the product’s voice by default rather than by review.
The practical measure: design review moved off pixels and onto behaviour and bacme discussions worth having.
- Define actions
- Build blocks
- Document behaviour
- Compose sessions
- Ship without redesign