AI Agent Engineer

AI Agent Engineer

Conditioning and Prompt Design for Image Generation

3 min read

Conditioning and Prompt Design

Goal of this lesson

Break a creative description into conditions a model can execute, a reviewer can check, and a version system can compare. A prompt is not a magic spell; it is part of a task specification.

A six-slot prompt structure

Fill the following slots from concrete to abstract:

  1. Purpose and canvas: e-commerce hero image, teaching illustration, or storyboard; ratio and safe area.
  2. Subject: who or what, number, appearance, and relative position.
  3. Action and relationships: what the subject is doing and how subjects interact.
  4. Environment and time: place, weather, era, and background complexity.
  5. Camera and light: shot size, viewpoint, focus, and key-light direction.
  6. Visual language: medium, palette, and texture; use describable features and avoid unnecessary imitation of living artists.

For example: “A 4:5 portrait image for the cover of a Chinese tutorial; a white robotic arm places a blue sample tube into a rack in a bright laboratory; medium shot, slightly top-down, subject centered low; cool-white key light with a little blue rim light; clean vector illustration; reserve the upper 25% as a clear title area; generate no text, logos, or people.”

Hard constraints, soft preferences, and acceptance

  • Hard constraints: must be satisfied or the candidate is rejected, such as subject count, prohibited logos, or title safe area.
  • Soft preferences: desirable but negotiable when in conflict, such as “warm but restrained.”
  • Acceptance criteria: how a person or program will judge, such as “both sample tubes are visible” and “no readable brand identifier.”

Do not treat a negative prompt as a universal filter. A model may not understand complex negation or may refuse a request before safety filtering. Put critical prohibitions both in pre-request policy checks and in post-generation human review.

Reference inputs and edit controls

A reference image can supply identity, layout, color, or style, but state which part is referenced and whether use rights exist. For a local edit, also record asset_id, source_revision, original-image/mask hash, allowed-change region, protected region, target description, and acceptance comparison. Object-level authorization/ACL must apply before any candidate scoring or model read. If a product supports structural-control maps or strength parameters, treat them as provider-specific capabilities: probe capability instead of hard-coding them into generic business logic.

Iterate instead of piling on words

Change only one major variable at a time: confirm layout, then subject, then style and text. For every round, retain prompt_version, reason for change, output ID, score, and the execution chain as transform_id. Assign release_id only when a candidate enters the publication queue; do not infer versions from filenames. If the same defect does not improve for several rounds, change task decomposition—for example, generate a text-free background first and add typography with a layout tool—instead of endlessly appending adjectives.

Common errors and diagnosis

  • Contradiction: “minimal background” while requiring ten props. Establish priority first.
  • Not observable: “more premium.” Rewrite it as visible palette, material, empty space, camera, and similar features.
  • Changing too much at once: you cannot know which change worked. Keep a single-variable experiment table.
  • Treating the prompt as a safety policy: enforce safety rules independently.

Exercise and self-check

Rewrite “make a beautiful AI poster” as a six-slot specification. Mark three hard constraints, three soft preferences, and four acceptance criteria. Then explain how you would decompose the workflow if text generation is unreliable.

Continue with Composition, Style, Text, and Editing.

References

Backlinks