Michael James OwusuCreative technologist

01 · Systems

Systems · AI image production

Content Studio Engine

A local studio and command-line engine for directing AI product photography. A brief and photographic recipe go in; linted, priced prompts produce takes; a person picks the winners; finals are versioned with their prompts, references and costs.

At a glance

Role

Designed the workflow and the director rules. Built the CLI engine and its web studio with Claude Code and Codex. Ran the shoots and made every selection.

Tools

  • TypeScript
  • React
  • Express
  • Vitest
  • Gemini 3 Pro Image
  • GPT Image 2

Outcomes

  • Produced the stills for the four campaigns on this site, with every run recording its prompts, references and cost.
  • A lint gate stops vague prompts before spend: a generic test prompt returned 13 errors in a dry run.
  • Re-runs only pay for what changed: unselected scenes are inherited from the previous version.

Links

Studio review page for Pip WASH: the scene prompt, the recorded reason for the winning take, and four takes at $0.134 each with the winner marked.
Review page, captured 15 September 2026. Four takes of the WASH packshot at $0.134 each; the winning take carries the reason it was chosen.

The problem

AI product images fail in familiar ways: invented or soft label text, packaging that drifts from shot to shot, plastic skin and flat, directionless light. Fixing those one prompt at a time is slow and hard to repeat on the next campaign.

I wanted to direct a shoot the way a photographer would, with the rules written down once and enforced every time, and with the cost of each decision visible before it is made.

How a shoot runs

  1. 01

    Brief and recipe

    Product facts, references and a photographic recipe: photographer, light, surface, grade and lens plan.

  2. 02

    Lint gate

    Prompts without a focal length, f-stop, light direction or light quality fail, as do generic terms such as "8k" or "stunning".

    Automatic check
  3. 03

    Priced dry run

    The model, image count and estimated cost are shown before any request is made.

    Priced
  4. 04

    Takes

    Each scene generates one or more takes on the chosen model and size.

  5. 05

    Pick the winner

    A person picks a take and records why. The reason is saved to the brand's taste notes.

    Human approval
  6. 06

    Anchor and inherit

    The scene 1 winner can anchor later scenes. Unchanged scenes are inherited from the last version at no cost.

  7. 07

    Versioned finals

    Each version stores its prompts, references, takes, winners and costs as JSON beside the images.

Human approval Priced Automatic check

Studio project editor for Pip WASH: brief, photographer, lighting, surface, colour grade and lens fields, the label sheet and approved packshot as references, and an estimate of about $0.54 for four images.
Project editor for the WASH packshot. The recipe fields feed every scene prompt, the label sheet and approved frame are the only references, and the estimate (four images, about $0.54) updates before Generate is pressed.

The lint gate in a dry run

Real CLI output from 15 September, run on a copy of the project folders. The first run is the Pip float campaign; the second is a deliberately generic test prompt. Dry runs make no API calls. Excerpt.

$ npm start -- --project pip-float --dry-run
ℹ Refs: product=4 model=0 style=1 location=0
ℹ [DRY RUN] Would generate 3 images on gemini-3-pro-image. Estimated cost: $0.72
 
$ npm start -- --project lint-demo --dry-run
✗ lint error scene 1 [no-lens]: no focal length (e.g. "85mm") — camera language is required
✗ lint error scene 1 [no-fstop]: no f-stop (e.g. "f/8") — aperture drives depth behavior
✗ lint error scene 1 [no-light-direction]: no light direction (camera-left/right, backlit, overhead, 45°…)
✗ lint error scene 1 [no-light-quality]: no light quality (soft / hard / diffused / raking…)
✗ lint error scene 1 [banned-generic]: "cinematic lighting" pulls toward the generic-AI average — name the specific setup/grade/photographer instead
✗ lint error scene 1 [banned-generic]: "8k" pulls toward the generic-AI average — name the specific setup/grade/photographer instead
[… 7 more errors and 2 warnings not shown]
ℹ [DRY RUN] 13 lint error(s) — a real run will refuse these prompts

Where the gate sits

src/modes/images.ts (excerpt)
const lintErrors = lintIssues.filter((i) => i.severity === 'error');
if (lintErrors.length > 0 && !options.skipLint && !options.dryRun) {
  throw new Error(
    `Prompt lint failed with ${lintErrors.length} error(s) — fix the prompts (see log) or pass --skip-lint`,
  );
}

if (options.dryRun) {
  const totalImages = selectedPlans.reduce((sum, p) => sum + p.takes, 0);
  const estimate = selectedPlans.reduce((sum, p) => sum + p.unitCost * p.takes, 0);
  logger.info(`[DRY RUN] Would generate ${totalImages} images on ${plans[0]?.model ?? '?'}. Estimated cost: $${estimate.toFixed(2)}`);
  return { version: 0, versionDir: '(dry-run)', scenes: [], totalCost: estimate /* … */ };
}

Lint runs on the director-written prompt before any spend. A real run stops on errors unless --skip-lint is passed on purpose; a dry run reports them and prices the run.

Review page for version 4 of the float campaign: scenes 1 and 2 inherited from version 3, and four new takes of scene 3.
Version 4 of the float campaign re-rolled scene 3 as four takes ($0.54). Scenes 1 and 2 were inherited from version 3. Take 3D shows a failure the rules guard against: the seed mark drawn as a droplet.

Model routing and prices

Model1K2K4KUsed for
Gemini 3 Pro Image$0.134$0.134$0.24Finals and most campaign work
Gemini 3.1 Flash Image$0.067$0.101$0.151Cheaper drafts
Gemini 3.1 Flash Lite Image$0.034Cheapest drafts

Per-image prices as recorded in src/lib/cost.ts, checked against Google's published pricing on 4 July 2026. Run files store the estimate for every image; invoices are separate.

Walkthrough of the local studio: dashboard, project editor and two review pages. Recorded 15 September 2026 with generation disabled. Content Studio Engine walkthrough, 16 seconds, silent · MP4, 1.0 MB

Where a person decides

  • Before spend

    The estimate is on screen before Generate is pressed, and a CLI dry run lists lint issues before any request.

  • Winner selection

    Every winner is picked by a person with a written reason, which feeds the brand's taste notes for later briefs.

  • Brand library

    Only takes a person tags are promoted into the brand library, which later runs draw on as references.

How it's built and tested

Tests
145 passing (Vitest, strict TypeScript): versioning, inheritance, caching, selection, lint rules and the local API. The CLI engine that made the campaign finals has 71 more.
Stack
TypeScript on Node 22, Express with server-sent events for live runs, React and Vite, @google/genai. Prompt enrichment speaks any OpenAI-compatible API.
Storage
No database. Each run is a folder of images with run.json and winners.json; a content-hash cache avoids paying twice for identical inputs.
Local by design
The studio server binds to 127.0.0.1 by default, so API keys and reference images stay on the machine.

Made with it

Two Pip WASH tubes on cobalt blue, one showing the front label and one the back with directions, ingredients and barcode.
WASH, front and back
Two Pip PRESS dropper bottles on coral, showing the front label and the back ingredients panel.
PRESS, front and back
Two Pip PULP jars on marigold yellow, showing the front label and the back ingredients panel.
PULP, front and back
Pip WASH, PRESS and PULP arranged on a lilac plinth against a lilac studio background.
Line-up on lilac

See the Pip campaign