← Guides · 日本語 · Português · 中文
On this page
- The pattern
- Step 1 — Brief like a creative director
- Step 2 — Fix the type system before the pixels
- Step 3 — Generate imagery inside the session
- Step 4 — Animate with mechanisms, not adjectives
- Step 5 — The critique loop is the craft
- Verify it
- Pitfalls
- Knowledge check
- Walkthrough — brief by mechanism, refine by critique, on Devin CLI v3000.14 (swe-2-flash-medium)
- Sources
Guide 8 — Design & visual work
Goal: use Devin as your art director and implementer for production visual design — generated imagery, animation systems, typography — through conversation alone. Level: intermediate Source deployment: the Spraycash landing page — a fintech site's entire visual system designed inside Devin chat with Kimi K3, every prompt published: hero imagery, layered cloud animations, SVG draw-on ring, three-font type system. Sources: full write-up with prompts
The pattern#
Design-via-agent fails when prompts describe vibes ("make it premium"). It works when you art-direct the way you'd brief a designer: references, constraints, and a critique loop. Devin's edge over a design tool is that it produces the implementation — the animation is code, the imagery is generated and placed, the type system is shipped CSS.
brief + references ──► Devin produces visual assets + code
▲ │
└──── critique loop: "the ring animation draws too fast,
slow the stroke-dashoffset easing to 1.4s" ──────┘
Step 1 — Brief like a creative director#
The Spraycash write-up's prompts share one shape — concrete nouns, no adjectives-of-taste:
Hero section: dark navy (#0A1628) background, a circular progress ring
as the visual anchor — SVG stroke-dashoffset draw-on animation over
1.4s ease-out, layered clouds behind it drifting on separate
translateX loops at 40s/60s/90s periods so the parallax never repeats.
Give it: palette hexes, the animation mechanism (not "make it animate"), timing values, layer order.
Step 2 — Fix the type system before the pixels#
Spraycash locked a three-font system early — display, body, mono — before touching components. Do the same; it's the single highest-leverage visual decision:
Type system: <display font> for hero + section heads, <body font> for
prose, <mono> for numbers/labels. Set the scale: 64/40/28/18/14.
Nothing else uses another size.
Step 3 — Generate imagery inside the session#
Devin generates and places assets — don't stock-photo a designed page. The write-up shows generated hero imagery matched to the brief's palette. Keep generated assets consistent by fixing the style directive once (palette, lighting, material) and reusing it for every asset request.
Step 4 — Animate with mechanisms, not adjectives#
| Vague (fails) | Mechanical (works) |
|---|---|
| "smooth entrance" | "stagger children 80ms, translateY(12px)→0, opacity 0→1, 400ms ease-out" |
| "subtle parallax" | "three layers, translateX loops at 40/60/90s periods" |
| "draw the ring" | "stroke-dasharray = circumference, animate dashoffset C→0 over 1.4s" |
Spraycash's ring, clouds, and staggered reveals were all specified at this level — and shipped as real CSS/SVG, not mockups.
Step 5 — The critique loop is the craft#
The write-up's real lesson: quality came from iterations — "the ring draws too fast," "clouds read as noise, drop opacity 30%," "mono font on the numbers." Budget 3–5 critique rounds per visual anchor. Devin holds the whole page in context; each round is a message, not a ticket.
Verify it#
- [ ] The page renders with zero stock imagery — all assets generated or coded
- [ ] Every animation is a named mechanism with timing values in the CSS
- [ ] One type system, one scale — no stray sizes
- [ ] Viewed at 3+ widths, nothing breaks or clips
Pitfalls#
- Adjective prompts. "Premium," "sleek," "modern" give slop. Mechanisms and values give craft.
- One-shotting. Accepting the first render is why agent-made pages look agent-made. The critique loop is where quality lives.
- Missing the lock-down. Set palette + type system as constraints first; assets generated before constraints are set never match after.
Knowledge check#
- Why did the Spraycash author publish every prompt verbatim? (The prompt sequence is the design artifact — art direction through conversation is reproducible only if the conversation is.)
- What does 'generated imagery + layered animation' require that a screenshot mock doesn't? (Shippable assets and a working animation system — the verification is a rendered page, not a picture of one.)
- When is chat-direction better than a design tool for landing pages? (When the output must be production code anyway — the design session is the implementation.)
Walkthrough — brief by mechanism, refine by critique, on Devin CLI v3000.14 (swe-2-flash-medium)#
A bare fintech page (index.html, default sans-serif), one Spraycash-style brief with real mechanisms — hex palette, stroke-dashoffset ring over 1.4s ease-out, three cloud layers at 40/60/90s, a 3-font scale — then one critique round with exact corrections. Both rounds happened inside one session.
Video (61 s, silent): bare page → mechanism brief → implementation → rendered hero → critique → refined render.
1. Before: three lines of copy and a button, nothing else
2. The brief's numbers landed verbatim in the code — plus a prefers-reduced-motion fallback it added itself
3. Round 1 output — real CSS/SVG, not a mockup
4. Critique is where the craft lives: exact values in, exact diff out — and it caught a mismatch we didn't mention
5. Round 2 output — the loop is a message, not a ticket
Sources#
- Designing the Spraycash UI with Kimi K3 in Devin — full prompts and iteration history
- Related: built-with-devin — patterns
