Ai Image Generation Rules

---
name: ai-image-generation-rules
description: The rules for AI-generated images in marketing — what to generate, what to screenshot, and what never to fake — plus the specs and placement discipline for featured images and infographics. Use when asked to generate images for a blog or ad, decide between a screenshot and a generated image, create infographics or hero images, work out why generated images feel off-brand or dishonest, set image policy for an agent, or someone says 'just generate a dashboard screenshot'. For the creative build around images, see static-ad-builder. For the QC gate, see ai-content-qc.
metadata:
  version: 1.0.0
---

# AI image generation rules

One rule sits above every other, straight from the course: **screenshots for real
interfaces and real data, generated images for concepts, frameworks and hero
art. Never fake a dashboard with an image model.**

A generated image is an illustration. The moment it pretends to be evidence — a
dashboard, a result, a review — it stops being design and becomes a fabricated
claim with pixels.

## Before you start

| Input | Why |
| --- | --- |
| What each image must do — decorate, explain, or prove | Proof is never generated; the job decides the tool |
| The brand lock — colours, fonts, style | Generated images follow the same lock as everything else |
| The image placeholders in the content | The writer marks where images go; generation fills, never invents slots |
| Whether real proof exists (screenshots, product photos) | Real always beats generated for its own job |

## Step 1 — Route each image by its job

| Job | Tool | Examples |
| --- | --- | --- |
| **Prove** — show a real interface, result, metric | Screenshot only | Ads Manager, GA4, SERP results, a live page, reviews |
| **Explain** — carry an idea | Generate | A framework diagram, a comparison, a decision tree |
| **Decorate** — hero and mood | Generate | Featured images, concept art |
| **Show the product** | Real photo/screenshot | Never a generated look-alike of your own product |

The course's screenshot skills exist precisely because proof has its own
pipeline — captions on captured proof even carry `Source: {domain}, captured
{date}` for honesty.

## Step 2 — Generate to spec, not to taste

The course specs for content images:

- **Featured image: 1200×675.**
- **Inline infographics: one or two per post, each carrying an actual idea** —
  a framework, a comparison, a decision tree. An infographic with no idea is
  decoration filed in the wrong column.
- Ad statics follow their own spec (1080×1350, brand lock) — see
  static-ad-builder.
- Respect the pipeline order: **images before publish, always** — placeholders
  shipping live is a known failure.

## Step 3 — Keep the brand lock

Generated does not mean exempt: colours from the brand hex, typography per the
lock when text appears, and no distortion of real product imagery composited in.
An off-lock image breaks batch consistency exactly like an off-lock ad.

## Step 4 — Caption and alt-text honestly

- Generated images are describable as what they are — a concept illustration.
  They never get captions implying capture ("our dashboard", "results from…").
- Every uploaded image gets alt text describing what it shows — the pipeline
  uploads with alt text as a standard step, not an afterthought.

## Output

```
# Image plan: <content piece>

| Slot | Job (prove/explain/decorate) | Tool | Spec | Status |
| featured | decorate | generate | 1200x675 | brand-lock ✓ |
| section 2 | explain (framework) | generate | inline | carries idea ✓ |
| section 4 | prove (SERP result) | screenshot | — | Source+date caption |

Proof images: <n> — all captured, none generated
Alt text: written for every slot
Order: images complete BEFORE publish
```

## When it breaks

| What you see | What it means | The fix |
| --- | --- | --- |
| A generated "dashboard" in the draft | The one rule broken | Replace with a real capture, or cut the claim it faked |
| Post shipped with IMAGE_PLACEHOLDER visible | Generation ran after publish | Pipeline order: images before publish, always |
| Infographics look nice, say nothing | Decoration filed as explanation | Each inline image carries a named idea or moves to hero duty |
| Images feel off-brand | Generated outside the lock | Brand hex and type rules apply to generated work too |
| Caption implies the image is real evidence | Honesty failure at the caption | Describe generated as concept; captured proof gets Source + date |
| Product images look subtly wrong | A generated look-alike of the real product | The product is photographed or screenshotted, never approximated |
| Accessibility flags missing alt text | Upload step skipped it | Alt text is part of the upload call, not a follow-up |

Never rescue a weak proof section by generating what the screenshot should have
shown. If the real capture is unavailable, the claim it supported waits — a
faked dashboard discovered once poisons every real screenshot you ever publish.

## Rules

- **Never fake a dashboard, result, or review with an image model**, because a
  generated image presented as evidence is a fabricated claim, and one discovery
  discredits all your real proof.
- **Proof is captured, concepts are generated** — the job decides the tool,
  because the two have different honesty contracts.
- **Images complete before publish**, because placeholders shipping live is a
  known, avoidable failure.
- **The brand lock covers generated images**, because "the AI made it" is not an
  exemption from being on-brand.
- **Every image carries honest alt text and captions**, because accessibility
  and honesty share the same field.

## Related skills

- **static-ad-builder** — ad-format image assembly under the same lock.
- **ai-content-qc** — the gate that catches a faked proof image before it ships.
- **env-and-keys-setup** — the image-model key and its pay-per-use billing.
- **design-system-spec** — where the brand lock's values live for web work.