Image generation with iterative refinement

Summary: An AI workflow that generates images based on a conversation about what you need. After discussing requirements, the AI generates an image, evaluates it against the requirements, identifies what needs improvement, regenerates, and repeats until it's satisfied the image meets the brief. Only then does it present the result for your review. You may still request changes, but the first version is already refined.

What makes this colleague-level: The AI works independently through multiple iterations, evaluating and improving its own output before involving you. Unlike automation (where AI does a task and you check it) or context (where you drive each step), this workflow delegates both execution and quality evaluation to the AI. It hands work back when it's confident you'll be satisfied — like a colleague who drafts, self-reviews, and only brings you the polished version.

Map

The recurring workflow is creating images for content, presentations, or marketing materials. It decomposes into: discuss requirements with the user (subject, style, mood, brand constraints), generate an initial image based on a structured specification, evaluate whether the image meets the stated requirements, identify specific improvements needed, regenerate with a refined specification, repeat the evaluation-refinement cycle until quality threshold is met, present the final result for human review.

The generation-evaluation-refinement loop is the core AI candidate — it benefits most from autonomous iteration because each cycle improves the output without requiring human input at every step.

Test

Initial test was a single-pass generation: describe what you want, get one image back. Results were inconsistent — sometimes excellent, sometimes missing key requirements entirely.

The breakthrough was adding self-evaluation: instead of presenting the first generation, the AI was instructed to check the image against the original requirements and regenerate if anything was missing or off. This single change — adding an evaluation step between generation and presentation — dramatically improved first-presentation quality. The AI caught issues (wrong color palette, missing elements, poor composition) that previously required human feedback loops to fix.

Integrate

Built as a skill with structured image specifications. The AI translates the conversation about requirements into a formal specification (layout, colors, content elements, style direction), then generates the image. After generation, it evaluates the result against the specification and either refines or presents.

The specification format ensures consistency across generations — the same type of request produces comparable quality each time. Brand defaults (colors, typography, visual style) are embedded in the skill so brand-consistent images don't require re-specifying brand guidelines with every request.

Compound

Two mechanisms make this workflow improve over time.

First, the specification templates improve. When a generated image requires human correction after the autonomous refinement cycle, the underlying specification format can be examined to determine whether a template change would prevent the issue. If images consistently need more whitespace, that preference becomes a default in the specification template rather than a correction applied each time.

Second, the retrospective loop: after the user approves a final image, the AI can analyze what changed between its initial specification and the final approved version. If the same type of adjustment happens repeatedly — the user always asks for bolder typography, or always requests a different crop — that pattern becomes a default in future specifications. This transforms individual corrections into systematic improvements to the generation process itself. Each image session makes the next one faster and closer to what the user wants on first presentation.

agent-design-principles — The self-evaluation loop embodies the "start with outcomes" principle — the agent knows what done looks like before it begins
compound-engineering — The retrospective loop is a concrete implementation of continuous improvement through use
productive-friction — The human still reviews the final output — full autonomy without review would remove valuable quality control