gpt-image2-prompt-director

Turn a weak idea into a high-impact GPT image2 (gpt-image-2) generation brief: prompt director workflows, idea generation, visual prompt frameworks, artifact specs for avatars, sticker/emoji packs, in

SKILL.md

GitHub

GPT Image2 Prompt Director

Create high-quality GPT image2 prompts from weak ideas, or evaluate generated prompts against the archived Xiaoxiaodong expert prompt benchmark.

Default Workflow

  1. Identify the input mode:
    • Weak idea: one theme, title, object, phrase, image concept, or product.
    • No idea: user only gives identity, platform, or goal; first generate candidate ideas.
    • Prompt repair: user provides an existing prompt; evaluate and revise it.
    • Evaluation: user asks to run or inspect the benchmark.
  2. Load references/prompt-framework.md when generating or repairing prompts.
  3. If the idea feels ordinary, avatar/IP-like, UI-like, or product-like, also load references/artifact-spec-framework.md. First choose the artifact container before writing the prompt.
  4. For avatar, profile-picture, sticker, emoji, meme, or expression-pack tasks, also load references/avatar-expression-framework.md. Treat the output as a reusable identity asset or expression system, not as a generic portrait.
  5. Generate one complete final prompt, plus 2-3 optional creative directions when useful.
  6. Before finalizing, check the prompt has:
    • role positioning
    • anti-definition
    • input contract
    • internal planning
    • content planning
    • visual system
    • anti-failure constraints
    • output spec
    • self-check
  7. If the user asks for validation, run the bundled evaluator.

No-Idea Mode

When the user has no idea:

  1. Ask for no clarification unless the goal is truly ambiguous.
  2. Infer a likely publishing context from the user request.
  3. Generate 10-20 candidate image plays.
  4. Score each by:
    • surprise
    • GPT image2 fit
    • visual clarity
    • shareability
    • ease of execution
  5. Pick the strongest direction and expand it into a complete prompt.

Wow Mode

When the first output feels generic or the user asks for something less ordinary:

  1. Stop optimizing adjectives.
  2. Pick a stronger artifact container.
  3. Decide whether a multi-cell grid/contact sheet would improve the result.
  4. Create 3-5 high-concept collisions.
  5. Prefer structured spec output for UI, social, avatar identity, poster systems, and product mockups.
  6. Add a concrete Definition of Done.

For avatar and personal-brand work, prefer an avatar identity system over a standalone portrait unless the user explicitly asks for a single image only.

Use n x n grid decomposition when variety is part of the value: avatars, stickers, meme packs, character variants, icon systems, material/style tests, storyboards, or repeated information cards. Prefer 3 x 3 for avatar exploration and 4 x 4 for sticker/expression packs. Avoid grids for cinematic posters, hero key visuals, luxury product shots, or any image that needs one dominant focal point.

Avatar / Sticker Rules

For avatar prompts:

  • Do not default to a cinematic portrait, fashion editorial, or realistic headshot.
  • If there is no reference image, prefer a 3 x 3 avatar exploration sheet first.
  • If the user needs one final avatar, make it a platform avatar asset: large head silhouette, simple background, one memorable symbol, one accent color, circular-crop safe, readable at 80px.
  • Keep the symbol secondary to the face. Do not cover most of the face with paper, crystals, UI, geometry, or decorative objects.
  • Prefer semi-realistic / semi-graphic / illustrated / badge-like treatments over photographic realism unless the user explicitly asks for a photo.
  • If a personal-brand avatar looks like a poster portrait, compare against examples/personal-brand-avatar-repair.md and repair toward avatar-asset mechanics.

For sticker and expression prompts:

  • Default to a 4 x 4 expression pack.
  • Use "consistent imperfection": the same character can be wrong, but must be wrong in the same way.
  • Text, when present, should feel hand-drawn with the image, not typeset afterward.
  • Each cell must be independently usable and emotionally clear.

Evaluation Commands

Create a runnable eval pack:

cd gpt-image2-prompt-director
node scripts/eval_prompt_director.mjs --init-run /tmp/gpt-image2-prompt-director-eval

Evaluate the archived expert prompts as a sanity baseline:

cd gpt-image2-prompt-director
node scripts/eval_prompt_director.mjs --gold --report /tmp/gpt-image2-prompt-director-gold-report.md

Evaluate generated prompts saved as outputs/<case-id>.md:

cd gpt-image2-prompt-director
node scripts/eval_prompt_director.mjs \
  --outputs /tmp/gpt-image2-prompt-director-eval/outputs \
  --report /tmp/gpt-image2-prompt-director-eval/report.md \
  --fail-under 80 \
  --strict

Evaluate one prompt:

cd gpt-image2-prompt-director
node scripts/eval_prompt_director.mjs \
  --case-id 09 \
  --prompt-file /absolute/path/to/prompt.md

For scoring details, read references/rubric.md.

Benchmark Data

The benchmark is in evals/benchmark-cases.json. It contains 40 cases derived from an archived prompt corpus and regression tests. Each case includes:

  • weak input
  • source title
  • category
  • expected capabilities

Use it to test whether a generated prompt reconstructs expert-level capability from weak input. Do not treat it as a memorization target.