name

design-image-studio

description

Generate design-quality AI images for posters, product visuals, PPT covers, infographics, and teaching diagrams. Use this when the user wants design-grade output rather than generic AI art — the skill compiles a loose brief into a structured design brief, picks the right Volcengine Seedream settings, and can generate the image directly. Also exposes an optional five-stage workflow (brainstorming → plan → generate → verify → iterate) for higher-stakes deliverables.

Design Image Studio

This skill turns a vague visual request into a usable design output. Not "AI art" — design output: posters, e-commerce hero shots, slide cover art, infographic-style visuals, teaching diagrams.

It runs in two modes depending on how high the stakes are:

Quick mode — user has a clear brief, just wants the image. Run scripts/design_image.py once, return.
Full workflow mode — user has a fuzzy brief or a high-stakes deliverable. Walk through brainstorming → plan → generate → verify → iterate.

The default is quick mode. Switch to full workflow when:

the brief is one sentence with no audience / no brand context, or
the user mentions "campaign", "客户给的", "正式上线", "印刷", "投放"

What Lives Where

Read these in order when doing serious design work:

references/claude-design-sys-prompt-full.txt — the upstream design brain (73KB, full Claude design-system prompt, do NOT collapse into bullet points)
references/claude-design-map.md — section index for the file above
references/design-compiler.md — how to compile the full prompt into design reasoning
The matching task reference: references/poster.md / product-image.md / ppt-visual.md / infographic.md / teaching-demo.md
If the first result is weak: references/anti-slop-and-failure-patterns.md

When to Use This Skill

Yes:

Posters / campaign key visuals / recruitment ads
Product hero shots / e-commerce main images
PPT cover art / chapter dividers / concept illustrations
Infographic-style visuals (background, not exact data charts)
Teaching/explanatory scenes
Visual concept exploration with stronger art direction than a generic image prompt

No:

Pixel-accurate UI recreation
Editable charts (use HTML/SVG/PPT)
Layouts that depend on rendered text accuracy (image models all struggle with Chinese small text)

Quick Mode

python scripts/design_image.py \
  --task poster \
  --brief "为 AI 训练营生成一张高冲击力招生海报，强调增长、实战和速度" \
  --direction balanced \
  --aspect 3:4 \
  --quality final \
  --output training-poster.png

If the user only wants prompts, add --prompt-only. The wrapper prints design_reasoning, compiled_brief, and the final prompt so you can verify the design system survived compilation.

Full Workflow Mode

Stage 1: brainstorming

Read BRAINSTORM.md. Ask the user the 5 core questions: usage, audience, brand anchor, must-include, must-avoid. Do not generate yet.

Stage 2: plan

python scripts/multi_direction.py \
  --task poster --brief "..." --aspect 3:4 --quality draft

Outputs outputs/multi-direction-<timestamp>/compare.md — a side-by-side of conservative / balanced / bold. Show this to the user; let them pick the direction.

Add --generate to actually produce three images for visual comparison (about ¥0.06 total in draft quality).

Stage 3: generate

After the user picks a direction:

python scripts/session.py init <session-name> --brief "..." --task poster
python scripts/design_image.py --task poster --brief "..." --direction <chosen> \
  --quality final --output outputs/<session-name>/v1.png
python scripts/session.py add <session-name> --version v1 \
  --image outputs/<session-name>/v1.png --note "first attempt"

Stage 4: verify

python scripts/verify_image.py \
  --image outputs/<session-name>/v1.png --task poster \
  --brand-color "#1A6FF7" --threshold 7.5 \
  --output outputs/<session-name>/v1.verify.md

7-dimension weighted score using doubao-vision. Exit code 3 means below threshold → iterate.

Stage 5: iterate

Read iterate_suggestions from the verify report. Apply deltas to --constraints / --avoid. Generate v2, verify v2, log v2 to session.

When v_n passes:

python scripts/session.py promote <session-name> --version v_n

Style Inheritance (optional advanced move)

If the user gives you a reference image (their existing brand asset, a competitor's hero, or "this kind of feeling"):

python scripts/style_extract.py --image refs/style-ref.jpg --as-cli-args

It prints ready-to-paste --constraints and --avoid snippets containing extracted dominant color, complexity, luminance, and (with --use-vision) semantic mood/composition descriptors. Splice them into your design_image.py call.

This is the right tool for batch consistency — five PPT covers in one deck, six product cards in one collection — share one style_token.json.

Things That Must Survive Compilation

When you write the prompt that goes to the image model, do NOT lose:

Purpose / audience / channel reasoning
Coherent visual system (one direction, not averaged styles)
One clear hero idea + supporting hierarchy
Whitespace and safe zones treated as design decisions
Anti-filler rules (every element earns its place)
Anti-slop rules (no random HUD overlays, no purple-blue gradient fog, no emoji-style decoration, no generic stock posing)
Brand context if provided

If you collapse the design system down to "premium, clean, modern" you've thrown away the value of the upstream prompt.

Execution Notes

Default final model: doubao-seedream-5-0-lite-260128 (best cost/quality)
Use --quality draft for direction exploration; switch to final only when direction is locked
For infographic / teaching tasks, do NOT ask the model to render dense small text — leave text-safe zones and overlay text in PPT/Figma later
The wrapper script is the compiler, not the design brain — never replace it with a few style adjectives
Session directory convention: outputs/<session-name>/v1.png + v1.verify.md + v1.brief.json + ... + final.png + manifest.json + CHANGELOG.md

Files

File	Role
`scripts/design_image.py`	Stage 3: design compiler + prompt builder
`scripts/generate.py`	Volcengine Seedream call wrapper
`scripts/multi_direction.py`	Stage 2: parallel three-direction explorer
`scripts/verify_image.py`	Stage 4: 7-dimension visual evaluator
`scripts/style_extract.py`	Style inheritance from reference image
`scripts/session.py`	Stage 5: session + CHANGELOG manager
`scripts/cli.py`	Unified `design-studio` subcommand entry
`references/claude-design-sys-prompt-full.txt`	Upstream design system prompt
`references/design-compiler.md`	Compilation workflow
`references/anti-slop-and-failure-patterns.md`	What to avoid
`BRAINSTORM.md`	Stage 1 question template
`WORKFLOW.md`	Full workflow documentation
`examples/case-*/`	Three end-to-end case studies