design

Files (1)

• SKILL.md

SKILL.md (first lines)

# design

Orchestrate visual design work. **You drive the tools, the navigator drives the aesthetics.**

This skill coordinates the full visual design workflow — from brand tokens to generated concepts to Figma vectors. It delegates to specialized skills (`/imagegen`, `/figma-conventions`, `/decision`) rather than duplicating them, and owns the transitions between phases.

The defining design principles:

> **The navigator is the art director.** Every visual decision — color, composition, mood, direction — is theirs. The agent proposes, generates, and implements; the navigator steers and approves. Never auto-advance past a visual checkpoint.
>
> **Brand tokens are the foundation.** Every design session starts from DESIGN.md. Colors, typography, and spacing are not invented per-session — they're read from the canonical source. Deviations are flagged, not silently applied.
>
> **Generate fast, implement slow.** Image generation is cheap (~2 min/variant). Figma vector work is expensive (~15-30 min/direction). Exhaust the ideation space with imagegen before committing to Figma implementation.
>
> **References, not assets.** Generated images are compositional references — layout, mood, tension. They are never imported into Figma as final artwork. The Figma implementation is built from components, variables, and vectors.
>
> **Eyes on every artifact.** The agent must actually view every design output — read generated images, capture Figma screenshots — and process the visual content before presenting to the navigator. Design review requires seeing, not just describing.

## Model quality for design work

Design tasks — especially visual review, composition analysis, and nuanced aesthetic judgment — benefit from higher-quality models. When the design skill is active:

1. **Detect the current model.** If the session is running on a fast/lite model (e.g. Haiku, Sonnet), note that design review quality may be limited.
2. **Suggest escalation for tricky work.** When the task involves reviewing generated images for compositional quality, comparing reference vs implementation, or making subtle aesthetic judgments, suggest the navigator switch to a higher-grade model:
   > "This comparison needs careful visual judgment. Consider switching to a higher-quality model (`/model opus` or similar) for this review phase — you can switch back after."
3. **Self-select when possible.** If the agent platform supports per-task model selection (e.g. spawning a subagent at a higher tier), prefer that over asking the navigator to reconfigure. The design review subagent runs at higher quality; the mechanical work stays on the current model.
4. **Don't nag.** Suggest once per session. If the navigator declines, proceed on the current model without further mention.

## Phase 0 — Load the design context

### Read DESIGN.md

Read `DESIGN.md` at the repo root. Extract and summarize the brand tokens the navigator will need:

- **Colors:** core palette (background, surface, text, accent, semantic)
- **Typography:** type ramp (families, sizes, weights)
- **Spacing:** if defined
- **Any brand-specific vocabulary** (logo shapes, motifs, naming conventions)

If `DESIGN.md` doesn't exist, warn:

> "No DESIGN.md found. I'll work without brand tokens — colors and typography will be freehand. Consider creating one via the DESIGN.md spec."

### Check imagegen prereqs

Verify `GEMINI_API_KEY` is available in the environment. If not, prompt per the imagegen skill's credential discovery flow. Don't block the session if the navigator wants to skip imagegen and go straight to Figma.

### Load the ticket (if invoked via /pair passthrough)

If this skill was invoked with a ticket ID (from `/pair` delegation or direct invocation), fetch the Linear issue and existing comments. Present a compact summary like pair does:

```
Design session: KB-67 — Design reusable pitch deck components
Status: In Progress | Priority: Low

Brand context (from DESIGN.md):
  Accent: #ff5583 | Background: #000000 | Text: #ffffff
  Type: Inter, 6-tier ramp (96px → 14px)

Prior work:
> [Summary from Linear comments if any]
```

If no ticket, this is a freeform design session — skip the ticket context.

### Present the phases

```
Design workflow:
  1. Concept  — generate visual candidates with imagegen
  2. Select   — review candidates, pick a direction
  3. Build    — implement the chosen direction in Figma
  4. Refine   — screenshot, compare, iterate

Start with concepting, or jump to a specific phase?
```

The navigator can enter at any phase. If they already have a direction in mind, skip straight to Build. If they have a Figma file in progress, skip to Refine.

## Phase 1 — Concept

### Build the prompt from brand context

When generating images, automatically weave in the brand tokens from DESIGN.md:

- **Always include:** the accent color hex, background color hex, and any named brand shapes (e.g. "the Asabina step shape — a hexagonal isometric polygon")
- **Add from the navigator's brief:** composition direction, style reference, mood, specific elements

Example prompt assembly:

```
Navigator says: "Brockmann-style poster with the step shape"

Assembled prompt:
"A Swiss poster-style graphic inspired by Josef Müller-Brockmann.
On a pure black (#000000) background, arrange isometric step shapes
— the Asabina brand element, a hexagonal polygon suggesting a 3D stair
step. One large saturated pink (#FF5583) step dominates, with smaller
dark gray (#2A2A2A) counterpoints. Hard geometric edges, flat opaque
color fills, strong asymmetric tension. No text, no gradients."
```

The agent enriches the navigator's brief with brand tokens — the navigator shouldn't have to remember hex codes.

### Generate candidates

For each prompt:

1. Write the prompt to `.imagegen-output/{name}.instruct.txt`
2. Call the Gemini API per the imagegen skill's API reference
3. Save the image to `.imagegen-output/{name}.png`
4. Auto-open on graphical systems (`open` on macOS, `xdg-open` on Linux)
5. Read the image inline for the navigator to review in the conversation

**Generate 2-3 variants per round** by default. Run the same prompt multiple times (image generation is stochastic) or vary the prompt slightly. Name them sequentially: `concept-v1.png`, `concept-v2.png`, `concept-v3.png`.

### Check in after each round

Present the candidates and ask:

From asabina-de/kb, CC-BY-4.0.

More from asabina-de/kb