Decision 0051 · Accepted
Design skill for
visual direction
Flow names design as agent work in the products built with it (PRODUCT.md) and keeps DESIGN.md as a root context file, but no skill owns choosing a visual direction.
- Status
- Accepted
- Built from:
- da88c48
Status: Accepted (2026-10-29). Spec: 039. Amended by 0067 (design is offered after the initial commit).
Context
Flow names design as agent work in the products built with it (PRODUCT.md) and keeps DESIGN.md as a root context file, but no skill owns choosing a visual direction. Adopter apps such as meta-cto (Signal) keep the template's look and a DESIGN.md with no identity. Visual choices are made ad hoc during build, or outside Flow, and are not recorded where the next agent looks. prototype covers in-app visual variants but stops at answering a question; it does not record an identity, and flow aligns DESIGN.md without choosing what goes in it.
Decision
Add design, a portable, optional skill for Flow-based apps ("design this"): read product and design context, state a short brief, render two to four distinct directions in context, recommend one, let the user choose, and record only the approved direction and its canonical assets in DESIGN.md on the work's branch, setting token values in the app's existing theme so record and code agree. Restyling components and pages stays with build. Neighbors keep their jobs: explore owns product discovery, prototype interaction and logic experiments (its UI guidance renders in-app variants for design), build applies it to components and pages, review checks rendered UI against DESIGN.md, and flow keeps DESIGN.md aligned. start stays fast: its baseline DESIGN.md says no identity is chosen yet, and it names design only as an optional step after the first deployment.
Rationale
Choosing a direction before building is cheaper than reworking surfaces after: rendered alternatives, a human choice, existing tokens respected, one source for shared assets, and old marks retired everywhere. A skill makes them reachable by phrase without a mandatory phase: no skill requires design, and work that fits an existing DESIGN.md skips it.
Alternatives
- Extend
prototype. Rejected: a prototype is disposable and answers one question; identity is durable and recorded. Mixing them would blurprototype's throwaway boundary. - Fold into
explore. Rejected: explore asks in proportion and ends in a product brief; visual exploration needs rendered alternatives and a separate approval. - Adopt a third-party design skill (for example Impeccable). Rejected as a whole: its command catalog and mandatory concept machinery exceed Flow's minimal process (research); its page-purpose and shared-versus-local principles are already in the methodology.
- No skill; document the practice only. Rejected: the practice already existed in the methodology and was not applied until the maintainer ran it by hand.
Consequences
The catalog grows to 26 skills. Templates pick design up on their next pnpm skills:update and should drop any third-party skill whose job it duplicates (0045).