AI workflow
Codex for marketing
Codex gets written off as a beginner tool. Put into a real marketing workflow — where the language model writes the prompts and the image model only renders them — it holds up. Here is where it fits, and where it does not.
Built from a public Techies Lab workspace: a launch campaign for Perplexity Computer, "The City Becomes a Computer." The workspace is an independent concept exercise — not affiliated with, authorised by, or endorsed by Perplexity or any brand shown.
Why it's relevant
Apply when: producing campaign creative at volume, evaluating whether an image model belongs in your stack, building a brand system an AI can actually use, or trying to work out why AI-generated visuals keep looking stiff and off-brand.
The one-line thesis
Most people fail with image models because they try to do both jobs in the image tool. Let the language model write the prompt and let the image model render it — the output stops looking generated and starts looking art-directed.
Who does what
| Step | Tool | Output |
|---|---|---|
| Strategy, concept, copy, image prompts | Claude Code | The campaign brief, creative platform, calendar, copy deck, visual briefs |
| Generating visuals from those prompts | Codex / ChatGPT image gen | The finished images |
| Assembling posts and the presentation site | Claude Code | Post mockups and a live page |
The division is the whole point. The image model is not asked to have the idea, choose the metaphor, or remember the brand. It is asked to render one scene that has already been decided.
The five steps
1 · Set the strategy. The concept, the emotional line, the palette and — critically — a do / avoid list are decided before a single image exists. Everything downstream serves this. Rewriting this one document plus swapping the reference boards is roughly 90% of what makes output feel on-brand for a different company.
2 · Write the prompts. Each asset gets a short, loose brief: a scene, a metaphor, and the exact words to render — not a rigid spec. This is the single biggest lever in the whole workflow.
3 · Generate. Open a fresh image session, attach the reference boards, paste the master prompt once, then send one brief at a time and ask for two variations. Steer with feel words — "warmer", "more grain", "softer UI", "calmer serif" — not with more rules.
4 · QA against the brand checklist. Every image is checked against the same four criteria before it survives: reflection rather than screen, soft natural palette, a human moment in frame, and text rendered verbatim. Misses go back to step 3.
5 · Assemble and ship. The finals go back to the language model, which lays them into per-platform post mockups and a presentation page.
The five rules that make it work
Separate the jobs. The language model writes prompts; the image model renders them.
References beat rules. Attach two or three real examples and say "match this world." A reference board carries more information than a paragraph of constraints.
Prompt the metaphor, not the feature. "A crystal ball holding a tiny glowing laptop" — not "show the AI agent."
Loose beats strict. Over-specified prompts read as stiff. Leave room and steer by feel.
Lock only the words. The exact copy is fixed; everything else is the model's to art-direct.
Where it does not hold up
Rendered text is unreliable. Letters mangle often enough that the practical answer is to regenerate the image with no text at all and set the type afterwards. Any workflow that depends on the image model getting a headline right will produce rework.
It will not carry the idea. Remove the strategy step and the same tool produces exactly the generic output it is dismissed for. The quality lives in the brief, not the renderer.
Consistency is bought, not given. Faces, products and typography drift between generations. Consistency comes from the reference boards and the QA pass, and both cost human time on every batch.
It is a production tool, not a judgement tool. Which angle deserves budget is still a human decision — the workflow only makes it cheap to have the options in front of you.
How it maps to our services
This is the shape of the work behind AI-native tooling: the repeated part of creative production is handed to a machine, and the brand context is written down once in a form the machine can actually use. The agent-ready brand system is the reference-board and do/avoid list, made durable.
The full workspace — briefs, creative platform, visual briefs and master prompt — is public: github.com/Techieslab-app/claude-codex-marketing-workflow.