ad-designer（广告设计师）

Ad Designer

Generate marketing ad images from campaign-planner creative briefs. Wrap nano-banana-pro with brand-aware prompt construction, aspect ratio handling, and a self-review loop that catches quality issues before marking images final.

Roles

• Ad Designer (this skill) — sole executor
• nano-banana-pro — image generation engine (upstream dependency)
• campaign-planner — upstream source of creative briefs
• brand-bible-builder — upstream source of brand visual identity

Prerequisites

Verify before starting:

command -v uv
test -n "$GEMINI_API_KEY"
test -f ~/.codex/skills/nano-banana-pro/scripts/generate_image.py

If uv is missing: tell the user to install it with curl -LsSf https://astral.sh/uv/install.sh | sh. If GEMINI_API_KEY is unset: tell the user to export it from Google AI Studio. If the script is missing: tell the user the nano-banana-pro skill must be installed first.

Step 1: Collect Inputs

Identify which inputs are present in the conversation or filesystem:

If the creative brief is missing, ask:

"I need a creative brief to generate the ad. Run /campaign-planner first, or paste the brief details: headline, visual direction, CTA, and target audience."

If the brand bible is missing, proceed with neutral visual defaults — note this in the output and recommend running /brand-bible-builder for on-brand results.

Step 2: Read Brand Bible Visual Identity

Open the brand bible and extract these fields into a working reference:

• Primary colors — hex codes or descriptors (e.g. #1A2E5A deep navy, #F5A623 amber)
• Secondary / accent colors — supporting palette
• Font style — serif vs sans-serif, weight, mood descriptor (e.g. "bold geometric sans", "light editorial serif")
• Imagery style — photography vs illustration, lifestyle vs product-only, color grading mood (warm/cool/muted/vibrant)
• Layout pattern — minimal and clean vs content-dense, whitespace-heavy vs layered
• Tone — formal, casual, bold, playful, premium, approachable

Keep this as a 6-field internal reference. Do not show it to the user. Use it when constructing prompts in Step 3.

Step 3: Construct Image Generation Prompt

Apply these rules every time. They are non-negotiable.

Core Rules (from best-practice image generation)

Rule 1 — Less is more. Short, focused prompts outperform long descriptive ones. Target 15–30 words for the visual description. Cut adjectives that repeat the same idea.

Rule 2 — No logos, brand names, or company names. Never include the brand name, product name, logo description, or any trademark in the prompt. The generation engine hallucinates these badly and they must be added in post-production.

Rule 3 — Only include text that is explicitly specified. If the brief specifies a headline to appear in the image, include it exactly as written — no paraphrasing. If no text is specified, do not add any text to the prompt. Do not include CTAs, taglines, or copy unless the brief explicitly says to render text.

Rule 4 — Describe emotion and composition, not data. Instead of "a woman saving money on her bills," write "confident woman, bright home office, warm morning light, relaxed expression." The emotion communicates the message; the model handles the rest.

Rule 5 — Translate brand colors without naming the brand. Convert hex codes to descriptors: #1A2E5A → "deep navy", #F5A623 → "warm amber". Use these color words in the prompt naturally.

Prompt Construction Formula

Assemble the prompt in this order:

[SUBJECT] — who or what is the focal element
[SETTING] — environment or background context
[MOOD / EMOTION] — feeling the image should evoke
[COMPOSITION] — framing, shot type, perspective
[LIGHTING] — quality and direction of light
[COLOR PALETTE] — 2–3 colors derived from brand bible
[STYLE] — photography/illustration style from brand bible
[AVOID] — explicit exclusions (text, logos, clutter, etc.)

Do not use all 8 fields every time. Use only those relevant to the brief. A 5-field prompt is often stronger than an 8-field one.

Example Construction

Brief input:

headline: "Finally, sleep that works"
visual_direction: "Woman waking up refreshed, natural light bedroom"
cta: "Try free tonight"
target_persona: "Aisha, 28, KL, stressed professional"

Brand bible: warm coral #E8735A, off-white #FAF7F2, editorial sans-serif, warm and minimal photography style.

Constructed prompt:

Young woman waking up peacefully, sunlit minimalist bedroom, linen textures,
warm coral and off-white tones, soft morning light through sheer curtains,
close-up on relaxed expression, warm editorial photography style

What was excluded: brand name, CTA text, product name, headline text (no text was specified for visual rendering).

Step 4: Select Aspect Ratio

Map the brief's platform or format to the correct ratio flag and pixel dimensions:

If the brief does not specify a platform, default to 1:1.

For carousel ads: generate all cards at the same ratio. Use 1:1 unless the brief specifies otherwise.

Pass the ratio to the script via --resolution. The nano-banana-pro script handles internal sizing. Generate at 1K for drafts, 4K for finals.

Step 5: Generate Images

Create the output directory:

mkdir -p /tmp/marketing/assets/images

Single image generation

Use this command pattern. Always use absolute path for the script.

uv run ~/.codex/skills/nano-banana-pro/scripts/generate_image.py \
  --prompt "<constructed prompt>" \
  --filename "/tmp/marketing/assets/images/<timestamp>-<brief-id>-draft.png" \
  --resolution 1K

Filename format: yyyy-mm-dd-hh-mm-ss-<brief-id>-draft.png

• brief-id: lowercase kebab-case derived from the brief headline (e.g. finally-sleep-works)
• Use current date/time for timestamp

Draft-first workflow

Always follow this sequence:

1. Draft at 1K — generate first at 1K for fast feedback. Present to user.
2. Iterate — if issues found in self-review (Step 6), adjust prompt and regenerate at 1K. Max 3 attempts.
3. Final at 4K — only after user approves the draft or after self-review passes. Change filename suffix from -draft to -final, set --resolution 4K.

Carousel ads

Generate each card as a separate image. Naming convention:

yyyy-mm-dd-hh-mm-ss-<brief-id>-card-01-draft.png
yyyy-mm-dd-hh-mm-ss-<brief-id>-card-02-draft.png

Ensure visual consistency across cards:

• Use the same color palette descriptor in every card's prompt
• Use the same style descriptor in every card's prompt
• Use the same lighting descriptor in every card's prompt
• Vary only the subject and composition per card

Step 6: Self-Review Loop (CRITICAL)

After generating each image, perform a structured self-review before presenting to the user. Do not skip this step.

Run through this checklist for every generated image:

Fail response

If any check fails:

1. Identify the specific failure from the table above.
2. Apply the corresponding fail action to the prompt.
3. Regenerate at 1K.
4. Re-run the self-review checklist on the new image.
5. Stop after 3 total attempts. If still failing on attempt 3, present the best result to the user with a note explaining what the issue is and what was tried.

Do not present a failing image silently. Always tell the user what was wrong and what fix was applied.

Step 7: HITL — Present for User Review

After self-review passes (or after 3 attempts), present the results:

Ad image(s) generated for: [brief headline]

Draft(s) saved to /tmp/marketing/assets/images/:
  [timestamp]-[brief-id]-draft.png   ← [ratio] | 1K

Self-review: [PASS / notes on any issues found and fixed]

Review the image and reply with one of:
  A) Approve → I'll upscale to 4K final
  B) Adjust → describe what to change
  C) Skip → move to next brief

Wait for the user's response before generating the 4K final or moving to the next brief.

If the user says "adjust" or describes changes:

• Apply the change to the prompt (keep changes minimal — one diff at a time)
• Regenerate at 1K
• Re-run self-review
• Present again

If the user approves:

• Regenerate at 4K with -final filename suffix
• Confirm: "Final saved: /tmp/marketing/assets/images/[filename]-final.png"

Step 8: Process Multiple Briefs

If the input contains multiple briefs (e.g. from a campaign-planner JSON with 3–5 image ad entries), process them sequentially. Complete the full cycle (generate → self-review → HITL approval → final) for each brief before moving to the next.

After all briefs are processed, present a summary:

All ad images complete.

Generated:
  /tmp/marketing/assets/images/[file-1]-final.png   [ratio]
  /tmp/marketing/assets/images/[file-2]-final.png   [ratio]
  /tmp/marketing/assets/images/[file-3]-final.png   [ratio]

Next step: run /meta-ads-publisher to upload these to your ad account,
or /campaign-planner to review the full campaign.

Error Handling

File Paths Reference

/tmp/marketing/
├── assets/
│   └── images/
│       ├── yyyy-mm-dd-hh-mm-ss-<brief-id>-draft.png    ← draft outputs
│       └── yyyy-mm-dd-hh-mm-ss-<brief-id>-final.png    ← approved 4K finals
├── brands/<brand>/
│   └── brand-bible.md                                   ← brand visual reference
└── campaigns/
    └── <brand>-campaign-plan.json                       ← input brief source