AI illustration system
for servers.com
Turning illustration from a per-asset cost into brand infrastructure: 200+ publish-ready visuals in one visual language, with no in‑house illustrator.
Animation built by the system for the streaming white paper
Challenge
Servers.com published across site, marketing, launches, talks and events with no in‑house illustrator. The routine was stock: search it, license it, redraw it. Hours an asset, and it never sounded like us.
Approach
I proposed a system where AI handles execution and all the value sits in the rules: six testable brand principles, a governed prompt template with negative prompts, and human decision gates in the pipeline.
Outcome
200+ illustrations in eight months, every one reading as servers.com without a logo anywhere in the frame. Illustration stopped being a bottleneck and became something the team can run without me.
Publish-ready illustrations shipped in eight months
Brief to approved asset, where stock search and redraw took hours
Bought-and-redrawn stock left the pipeline for good
Visual language across site, campaigns, decks, events
Output is the published set, counted from the library; speed compares a brief-to-approved run today against searching, licensing and redrawing stock.
Every frame is the output of rules I can defend
The conventional routes didn't add up, so I moved the value upstream. AI executes; the designer decides. Palette, light, abstraction level, motif, composition and bans are fixed before anything is generated, which is what makes the result reproducible by anyone on the team, at any volume.
The visual language that gates every generation
From one-off pictures
to a production pipeline.
Define
Lock what a servers.com illustration must be — six testable principles,
not a mood board.
Encode
Turn the principles into a prompt template, negative prompts and explicit reject criteria.
Delegate
Ship it as a configured agent so the brand language holds when someone else operates it.
Govern
Human gates at concept, brand comparison and final approval.
No gate passed, no publish.
Same three layers — only the engines under them changed
Built in 2025. Here is what I would build today.
In 2025 this ran on ChatGPT / DALL·E and Midjourney, because that was the reliable toolset for governed enterprise output — reference conditioning, style locking and open pipelines were not there yet. The part that lasts is the architecture: rules first, execution delegated, human gates where quality is decided. Today I build the same system differently — reference-conditioned models and a brand-trained LoRA for style locking, a node-based or API pipeline for batch runs and reproducible seeds, and MCP-connected agents wired straight into the content calendar. I run those workflows now; this case is the thinking that makes them safe to hand over.
Six rules that hold the style
Palette derives from the brand. Light is directional, soft and warm. Composition is asymmetric, with air. Abstraction sits mid-scale — volumetric, never photoreal, never flat. Motifs stay infrastructural. Bans are absolute: no faces, no text, no icons,
no competitor products.
Six rules that make the output reproducible.
These six rules are the interface to the system. Any generation can be accepted or rejected against them in under a minute, by me or by anyone else on the team.
- Palette derives from the brand, never from outside it
- Light is directional, soft and warm — quiet power
- Composition stays asymmetric and leaves air in the frame
- Volumetric geometry at mid abstraction, never photoreal
- Motifs stay infrastructural: racks, nodes, planes, cores
- Hard bans: no faces, no text, no icons, no competitor products
Blog and white-paper covers generated by the system — one visual language, zero hand-drawn illustration
The system ships as an agent
Instructions carry the brand context, the prompt template and the rejection criteria; the brand manual pages sit in the agent's knowledge base. The system stopped being my personal trick and became a shared asset — the visual language holds even when I am not the one operating it.
The actual agent configuration that ships these illustrations
Workflow diagram from the brand manual — mandatory steps, owners and the gates that cannot be skipped
Everything a teammate needs
to run it unsupervised.
The real setup: instructions, negative prompts, capabilities and conversation starters — everything a teammate needs to run the system unsupervised.
Every rule that governs the style is written down and shipped with the agent.
Illustration became infrastructure.
The value moved from producing pictures to owning the language behind them. I made fewer illustrations by hand than ever before, and held the brand's visual language tighter than ever before.
- 200+ publish-ready illustrations in eight months
- Same-day turnaround from brief to approved asset
- One visual language on site, in marketing and at events
- A system the team runs themselves
If this is the kind of case you are hiring for, I am happy to walk through the decisions behind it.