30x-image

image_generation

• A library of

  examples/<brand>/DESIGN.md

  files containing brand profiles (VoltAgent 9-section format + YAML frontmatter token block +

  taste:

  block).
• A

  references/

  library with: 7-template prompt taxonomy, combinatorial variation axes per template, and anti-slop banned rules.
• The Codex built-in

  image_generation

  tool — no API key needed.

30x-image

1. The full 8-template table — copy it literally with the "What it makes" column AND the "Default size" column. Bare template names (e.g. just

   logo, ad-creative, slide, ...

   ) are NOT acceptable — users need to see what each makes to choose.
2. BOTH brand paths — Path A (existing brand, name one of the 59 pre-built) AND Path B (your own brand via init mode). Never show only Path A. Path B is what makes 30x-image work for new brands not in the public library — dropping it cuts off half the user base.
3. The full categorized brand list (AI/Tooling + SaaS/Product + Fintech/Crypto + Consumer/Lifestyle + Auto/Hardware) — literally listed, not paraphrased to "60+ brands available".
4. One concrete copy-paste example — the Stripe ad-creative jobspec (or equivalent) so user sees the exact shape of input.

First time? Just try a popular brand.

npx getdesign list

ships with 60+ ready-to-use brand profiles — Stripe, Linear, Notion, Apple, Tesla, Nike, Spotify, Starbucks, Figma, Cursor, Claude, and more. Pick any one you recognize, run a 5-minute test (e.g.

template: ad-creative

with that brand), and see what 30x-image actually does on a brand you already have a feel for. Way faster than describing your own brand from scratch on the first try.

logo

ad-creative

slide

product-mockup

marketing-with-text

lighting-transform

scene-with-person

carousel

npx getdesign list

npx getdesign add <brand>

template:

subject:

copy:

taste:

Use 30x-image. Generate an ad-creative for Stripe.

DESIGN.md path: ~/.agents/skills/30x-image/examples/awesome-vendor/stripe/DESIGN.md

Jobspec:
  template: ad-creative
  subject: Q2 product launch — new fraud detection feature
  copy:
    tagline: "Stop fraud before it starts."
    cta: "Try it free"
  size: 1024x1536
  quality: high
  n: 4

1. Which template? (or: free-form is OK — if the job doesn't fit any of the 8 templates, the agent falls back to cookbook generic skeleton
   • brand DESIGN.md + anti-slop, with no axis commitments. Slightly less deterministic but valid.)
2. Which brand? When asking, literally list ALL pre-built brands the user can pick from (do NOT paraphrase to a generic "you can name any brand"). Run

   npx getdesign list

   to get the live full set, OR use this confirmed snapshot of 59 brands (verify with

   npx getdesign list

   in case the upstream library expanded):

   AI / Tooling: Claude, Cohere, Composio, Cursor, ElevenLabs, Expo, HashiCorp, Lovable, MiniMax, Mistral.ai, NVIDIA, Ollama, OpenCode.ai, PostHog, Raycast, Replicate, Resend, RunwayML, Superhuman, Together.ai

   SaaS / Product: Airbnb, Airtable, Cal, ClickHouse, Figma, Framer, IBM, Intercom, Linear, MongoDB, Mintlify, Miro, Notion, Sanity, Sentry, Shopify, Stripe, Supabase

   Fintech / Crypto: Binance, Coinbase, Kraken, Mastercard, Revolut

   Consumer / Lifestyle: Apple, Meta, Nike, Pinterest, PlayStation, Spotify, Starbucks, theVerge, Uber

   Auto / Hardware: BMW, Bugatti, Ferrari, Lamborghini, Renault, SpaceX, Tesla

   Just name any one — agent runs

   npx getdesign add <brand>

   and pulls a ready DESIGN.md in ~30 seconds.

   Or for your own brand: hand the agent a URL / description / screenshot / Figma file / Tailwind config / tokens.json / CSS — agent runs init mode (Mode 1) and synthesizes a DESIGN.md.
3. What's the brief? Subject + tagline + optional CTA. Or for carousel: per-slide subject/copy. Or for

   lighting-transform

   /

   scene-with-person

   /

   product-mockup

   : an input image to edit.

npx getdesign add <brand>

examples/awesome-vendor/stripe/DESIGN.md

## 8. Responsive Behavior

• $HOME/.30x-image/profiles/<brand-slug>/DESIGN.md

  if profile dir exists
• otherwise

  ./DESIGN.md

  in cwd

1. npx getdesign add <brand>

   — if user's brand is in the 60+ public library (Stripe / Linear / Notion / Apple / Tesla / Nike / Spotify / Starbucks / etc.), pull the ready DESIGN.md. This is the highest quality path — use it whenever possible. Augment with a

   taste:

   block (see Step 3).
2. Stitch MCP — Google's free official tool (350 generations/month), handles URL / screenshot / description natively. Setup:
   • User generates an API Key from Stitch Settings (https://stitch.withgoogle.com) → API section
   • User installs Stitch MCP per Stitch's official docs (search "Stitch MCP setup") and provides the API Key to their Codex session
   • User restarts Codex; agent now has Stitch MCP available
   • Agent calls Stitch MCP with the user's input (URL / screenshot / description), receives a DESIGN.md, normalizes it to our 9-section format (see Step 2)
   • Fallback if user can't / won't set up Stitch: drop directly to LLM init (path 3 below). Don't block on Stitch.
3. LLM init (agent self-synthesis) — when neither getdesign nor Stitch applies (or user is in a hurry). Agent reads the input source and writes a DESIGN.md from scratch following the Stripe DESIGN.md shape:
   • URL → fetch the page; extract palette from CSS /

     <meta theme-color>

     ; extract typography from

     font-family

     declarations; take a screenshot and visually verify atmosphere; write 9 sections of prose
   • Description → LLM-only inference; lower confidence; mandatory

     # auto-inferred, please review

     markers on every value
   • Screenshot only → vision-LLM extracts colors + typography vibes
   • Tailwind config / tokens.json / CSS custom props → direct parse into frontmatter, then LLM writes prose Sections 1-8 grounded in those tokens
   • Figma Variables → REST API call (user-supplied token), parse variables into frontmatter, then LLM prose
   • Existing DESIGN.md (Google format / external source) → convert section names to our 9-section format, augment with

     taste:

     block

---
version: alpha
name: <brand>
description: <one-line brand positioning>
colors:
  primary: <hex>
  secondary: <hex>
  ...
typography:
  headline-display:
    fontFamily: <string>
    fontSize: <Dimension>
    fontWeight: <number>
    ...
rounded: { sm: ..., md: ..., lg: ..., full: ... }
spacing: { xs: ..., sm: ..., md: ..., lg: ..., xl: ... }
components:
  button-primary: { ... }
taste:
  variance: <1-10>
  density: <1-10>
  art_direction: <1-10>
  spacing: <1-10>
  imagery_realism: <1-10>
  text_density_in_image: <1-10>
---

• DESIGN.md path: required. May be in cwd, in

  examples/

  , or

  $HOME/.30x-image/profiles/<brand>/DESIGN.md

  . Resolution priority:
  1. User-explicit path
  2. cwd

     ./DESIGN.md

  3. $HOME/.30x-image/active

     → that profile's DESIGN.md
  4. Otherwise prompt user to choose
• Template: one of

  logo / ad-creative / slide / product-mockup / marketing-with-text / lighting-transform / scene-with-person / carousel

  (carousel is multi-slide: see Step 4·B below). Resolution priority:
  1. User explicitly names a template → use it directly.
  2. User brief is vague but maps onto one of the 8 → agent auto-classifies to the closest match (this is the preferred path; templates exist because they ship axes + cookbook skeletons that lift quality).
  3. Brief is genuinely outside all 8 templates (e.g. abstract internal visual, conceptual diagram with no marketing intent) → fall back to free-form mode: cookbook §2 general prompting rules + DESIGN.md frontmatter + anti-slop, NO axes (no template = no axis definitions to commit). Free-form is fallback only — never the default when a template fits.
• Jobspec: subject + copy.tagline + copy.cta + size + quality + n + mode (override defaults from

  references/cookbook-7-templates.md

  )

taste:

• ## 1. Visual Theme & Atmosphere

  (gives image atmosphere)

• ## 2. Color Palette & Roles

  (every color name + hex + role)

• ## 3. Typography Rules

  (especially the typography table)

• ## 7. Do's and Don'ts

  → the Don't list is the brand's anti-slop banned

• ## 9. Agent Prompt Guide

  → if it has a

  ### For: 30x-image generate <template>

  section for the requested template, prefer that brand-specific cheat sheet over the generic

  references/cookbook-7-templates.md

  skeleton.

references/combinatorial-axes.md

taste:

• Higher

  variance

  → asymmetric / artistic options
• Higher

  density

  → packed / dense options
• Higher

  imagery_realism

  → photoreal / hard-light options
• Higher

  art_direction

  → bold / experimental options Commit the picks: write them into the prompt explicitly. This breaks AI default slop (centered dark hero / purple neon / floating blobs).

[Template skeleton from cookbook-7-templates.md OR Section 9 brand cheat sheet]

VISUAL ATMOSPHERE:
[paraphrase Section 1 essence — 2-3 sentences]

PHOTOGRAPHY / ILLUSTRATION STYLE:
[derive from taste.imagery_realism + Section 1 prose]

COMPOSITION:
- Axis 1: [committed choice]
- Axis 2: [committed choice]
- Axis 3: [committed choice]
- Axis 4: [committed choice if template has 4]

PALETTE (use ONLY these tokens, do not fabricate new colors):
- Primary: {colors.primary} ({brand-color-name from Section 2})
- Surface: {colors.surface}
- Headlines: {colors.on-surface}
- Body: {colors.neutral}
- Accent: {colors.tertiary} (sparingly)

TYPOGRAPHY IN IMAGE:
- Display: {typography.headline-display.fontFamily} weight {fontWeight}
- Letter-spacing: {letterSpacing}
- Font features: {fontFeature if set}

SUBJECT / COPY (from jobspec):
- Subject: {jobspec.subject}
- Tagline: "{jobspec.copy.tagline}" (verbatim, in quotes)
- CTA: {jobspec.copy.cta if provided}

ANTI-SLOP CONSTRAINTS (always include — read references/anti-slop-rules.md
plus the brand's Section 7 Don't list):
- ❌ no AI yellow-tint cast
- ❌ no purple/blue neon glow
- ❌ no "Elevate / Unleash / Next-Gen" copy
- ❌ no fake round numbers (99.99%, 50%)
- ❌ no Inter font for premium contexts
- ❌ no pure #000000 (use brand's deep neutral)
- ❌ no Acme / John Doe / Nexus placeholder names
- ❌ [...add brand-specific rules from Section 7 Don't]

image_generation

references/cookbook-7-templates.md

• size

  : e.g. 1024x1536 for ad-creative

• quality

  : low for explore, high for final

• n

  : 4 for explore, 1 for final

• action

  : "generate" for new, "edit" for iterate

• output_format

  : png

image_generation

1. Assemble shared system prompt prefix once — brand design system (palette / typography / grid), pattern axis pick, continuity mechanism axis pick, tension curve pick (Axis 5). This prefix is identical across all slides.
2. Plan slide roles AND narrative beats:
   • Slide 1 = cover (use Cover Slide Style axis).
   • Slides 2..(N-1) = body (use Body Slide Rhythm axis).
   • Slide N = CTA.
   • Apply tension curve (Axis 5) to assign each body slide a narrative beat: e.g. for

     Build-up

     , body slides escalate stakes 2→N-1; for

     Reveal

     , slides 2-3 withhold, slides 4..(N-1) reveal progressively; for

     Linear arc

     , slide 2=context, 3=challenge, 4-5=solution, N-1=result.
   • If user supplied per-slide

     copy

     : respect their order; only overlay tension curve as visual emphasis (e.g. slide N-1 gets larger headline, more saturated color — the "payoff slide").
   • If user only supplied

     subject

     and let agent write copy: agent MUST generate copy that satisfies the tension curve obligations. Sanity-check: does slide N-1 feel like the payoff? Does slide 1 hook toward it?
3. Fan out ALL slides in parallel as independent

   action=generate

   calls — same logic as every other template, just N calls instead of 1:
   • Build per-slide prompt = shared design-system prefix +

     THIS SLIDE: slide N of n_slides, role: <cover|body-N|cta>, subject: <jobspec.slides[N].subject>, copy: "<verbatim>"

     .
   • Every slide:

     action=generate

     ,

     n=1

     , quality at user's level (default

     high

     ). No

     action=edit

     chain. Reason: editing slide 1 repeatedly can trap gpt-image-2 in "webpage layout" mode and make slides look CSS-rendered rather than gpt-image-2-generated. Each slide must be a standalone generation, exactly like ad-creative or logo.
   • Cross-slide visual consistency comes from the shared prompt prefix (brand palette / typography / Visual Continuity Mechanism axis pick), NOT from image-reference chains. Trust the prefix.
   • Fire all N calls in parallel — wall-clock = ~1 image_generation call.
   • If a body slide drifts off-brand vs siblings, fix by reinforcing the shared prefix for that slide (more specific palette / typography directives), not by switching to action=edit.
4. Save each slide to

   candidates/slide-{N}.png

   (not

   1.png

   /

   2.png

   — slide-numbered for clarity).
5. manifest.json records:

   n_slides

   ,

   pattern

   ,

   slide_roles[]

   , full prompt per slide, hash of slide 1 (continuity reference).

$HOME/.30x-image/profiles/<brand-slug>/marketing/<job-id>/

• candidates/{1..n}.png

  (each candidate)

• manifest.json

  (see structure below)

examples/

./marketing/<job-id>/

{
  "job_id": "ISO-timestamp_short-slug",
  "design_md_path": "...",
  "design_md_hash": "sha256:...",
  "template": "ad-creative",
  "jobspec": { ... },
  "axes_picks": {
    "Hero Composition": "Asymmetric Split",
    "Subject Framing": "Medium environmental",
    "Lighting Mood": "Soft natural daylight",
    "Text-Image Integration": "Tagline overlaid bottom-left"
  },
  "tool_params": { "size": "1024x1536", "quality": "high", "n": 4 },
  "prompt_full": "...",
  "outputs": [
    { "path": "candidates/1.png", "selected": false },
    { "path": "candidates/2.png", "selected": false },
    ...
  ],
  "created_at": "..."
}

• Where the candidates are
• Which axis picks were committed (transparency)
• Optional: suggest "select one and run again with action=edit, quality=high" for the final pass.

1. Primary color from frontmatter is visibly dominant (CTA / accent)
2. Typography style matches

   ## 3. Typography Rules

   (e.g., Stripe = light weight 300, NOT bold)
3. No item from anti-slop banned list violated
4. Mood / density / variance match the

   taste:

   block

• DESIGN.md missing → 3-step fallback (try in order, tell the user which one you're taking):

  1. npx getdesign list

     — check if the brand is in the public awesome-design-md library (60+ brands: Stripe / Linear / Notion / Vercel / Apple / Tesla / Nike / Spotify / Starbucks / Apple / Coinbase / etc.). If yes →

  2. npx getdesign add <brand>

     — pulls a ready DESIGN.md to cwd in one command. Use it directly; optionally augment with a

     taste:

     block (5-7 numeric brand parameters) for axis biasing.
  3. Brand not in library → run

     init

     mode (M2) or hand-write a DESIGN.md using the 9-section format + frontmatter token block +

     taste:

     block (see

     examples/awesome-vendor/stripe/DESIGN.md

     as the canonical reference).
• Template name wrong → list 8 valid templates (incl. carousel)
• User brief doesn't match any template after honest attempt to classify → enter free-form fallback (NOT refuse). Tell user "no template matched, using free-form with brand DESIGN.md + anti-slop only" so they know determinism is reduced this run.
• Jobspec missing required fields (subject) → ask before invoking tool
• Brand

  ## 7 Don't

  is empty/weak → fall back to generic

  references/anti-slop-rules.md