Images

Decision

Use Performance Lab as the design and communication foundation for CREATE SOMETHING images: calm hierarchy, decisive contrast, plain claims, compact proof, governed execution, visible customer evidence, readiness states, and restrained action states.

CREATE SOMETHING owns the implementation language. Our images should show how AI work gets mapped, integrated, governed, validated, shipped, and handed off. For external education, marketing, social, deck, and client-proof surfaces, the approved OpenAI generated export is the first-class visual when it communicates better than a raw SVG or Atlas render.

Use docs/IMAGE_LANGUAGE_FOUNDATION.md when a generated, designed, captured, or rendered image needs TASTE-backed judgment context. TASTE references are inputs for review and prompting; they are not source assets to copy.

What images should prove

Every generated or designed image must answer at least one operational question:

1. What object, workflow, or system is being mapped?
2. What can run, needs review, or must stop?
3. What policy, contract, receipt, trace, eval, or screenshot proves the claim?
4. Who owns the next step?

If the image cannot answer one of those questions, it is decoration. Do not use it.

Visual grammar

Use this foundation for marketing materials, article visuals, social cards, pitch decks, and client updates:

Layer	Rule
Canvas	Prefer porcelain, white, or quiet near-black surfaces. Avoid noisy gradients.
Hierarchy	One plain claim, then proof beside or below it.
Structure	Use maps, lanes, cards, state rows, gates, receipts, and arrows.
Brand mark	Use the isometric cube as the persistent system signature.
Color	Use Canon clear tokens first: onyx, porcelain, cobalt, moss, and stop red.
Type	Use direct labels. Prefer workflow nouns over category language.
Motion	For video or animated exports, move only state, selection, progression, or handoff.
Proof	Make validation visible through receipts, tests, links, dates, owners, and status.

Raw SVGs, Atlas renders, and graph exports are useful control artifacts. Publish them only when they meet the same quality bar as a generated export; otherwise use them as source briefs for OpenAI generation or edits.

Image families

Create reusable image families instead of one-off illustrations:

• System map hero: the full operating path from input to governed outcome.
• Database / Automation / Judgment diagram: the three-tier framework as lanes or columns.
• Policy gate chart: allowed, ask, blocked, escalated, complete.
• Evidence map: artifact cards connected to the claim they prove.
• Handoff receipt: owner, state, validation, rollback, and next action.
• Screenshot annotation: real product evidence with restrained callouts.
• CTA visual: the specific next action, not generic AI promise.

GPT Image 2 prompt contract

Use OpenAI gpt-image-2 for production image generation when access is available. Keep prompts structured and repeatable:

Model: gpt-image-2
Quality: high
Size: <target size>

Create a CREATE SOMETHING <image family> for <surface>.
Purpose: <what the image must prove>.
Audience: <operator, buyer, builder, reviewer, client>.
Show: <workflow objects, states, proof artifacts, owners, gates>.
Style: CREATE SOMETHING Performance Lab image language:
plain claim, decisive hierarchy, compact proof, governed execution, visible
evidence, readiness states, and restrained action states. Translate that into CREATE SOMETHING
artifacts: porcelain `#f9f9f9`, white panels, onyx `#0a0e19`, quiet grey
`#636363`, thin `#e1e1e1` borders, 4-8px radii, compact receipts, state rows,
decision panels, artifact cards, policy gates, owners, proof rails, and a small
isometric cube signature only when useful. Use ocean `#0048ff`, moss `#1e3c2c`,
and stop `#c41e3a` only as semantic state accents.
Avoid: glowing robots, circuit faces, blue AI gradients, generic brains,
fake dashboards, stock photography, generic premium desk scenes, unreadable file
paths, private data, client secrets, vendor endorsement, watermarks.

The prompt should name the artifact family and proof requirement before style. Style cannot rescue a vague visual brief.

For article, social, deck, sales, or client-update images, start from packages/agency/content/templates/marketing/image-prompt.md and store the completed prompt beside the generated export.

Use the Image API for a single completed generation or edit. Use the Responses API image generation tool for conversational, multi-turn image refinement. Prefer editing from an approved export for revisions instead of regenerating from scratch when layout continuity matters.

Review gate

Before publishing or reusing an image:

• Text is legible at 50% size.
• The image still works without animation.
• The claim is supported by visible proof or a real screenshot.
• No fake UI is presented as a screenshot.
• No secrets, private data, private prompts, client records, or tokens appear.
• The asset has source prompt, source brief, model, date, output hash, owner, target surface, and refresh date.
• The result feels like governed operations, not generic AI atmosphere.

Use deterministic repo checks as the required gate for this workflow. Langfuse is not required to generate, store, or approve image assets. Add Langfuse later only if CREATE SOMETHING needs a scored rubric across many generated images, prompt variants, or approval outcomes.

File metadata

Every generated image should keep a source prompt beside the export:

<asset-slug>--prompt--vYYYYMMDD.txt
<asset-slug>--brief--vYYYYMMDD.md
<asset-slug>--export--1200x630--vYYYYMMDD.png

Include the model, snapshot if known, quality, size, source inputs, output hash, and review status in the prompt file. This makes the visual system auditable and repeatable. For article image sets, copy packages/agency/content/templates/marketing/image-metadata.md into the asset folder and fill it before publish.

Run the metadata/template check before publishing reusable article assets:

node scripts/marketing-image-assets-check.mjs

Boundary

Do not copy third-party identity, campaign language, page layouts, or category framing. Use CREATE SOMETHING Performance Lab as the standard for how governed autonomy should communicate through real systems, policy, evidence, and delivery.

Do not route Canon image generation through Google Gemini, Omni, or Vertex. This is an intentional stack-limit decision, not a temporary account-availability fallback. If OpenAI image access is blocked, keep the prompt and source brief ready rather than silently downgrading to another provider.