agentleFS
hanzhangzzzhanzhangzzz/agent-skills-zh/gpt-image2-prompt-director/SKILL.md

gpt-image2-prompt-director

Skill1 starsChanged 18 days agoMITView on GitHub

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, infographics, platform covers, posters and product shots, plus benchmark evaluation against an expert prompt set. Use when the user wants to create, improve, repair or systematically evaluate image generation prompts, or says $gpt-image2-prompt-director, 生图提示词, 图片提示词优化, 头像提示词, 表情包, 信息图提示词, 海报提示词, 修一下这个 prompt.

The file

Show its rules and text

We split this skill into the rules it gives an agent

19 rules. About 0 are common, 16 are distinctive, and 3 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈16

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 3

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

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

    line 10
  2. Distinctive:

    If the idea feels ordinary, avatar/IP-like, UI-like, or product-like, also load references/artifact-spec-framework.md.

    line 20
  3. Distinctive:

    Before finalizing, check the prompt has:

    line 23
  4. Distinctive:

    If the user asks for validation, run the bundled evaluator.

    line 33
  5. Distinctive:

    When the user has no idea:

    line 37
  6. Distinctive:

    Ask for no clarification unless the goal is truly ambiguous.

    line 39
  7. Distinctive:

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

    line 52
  8. Distinctive:

    Prefer structured spec output for UI, social, avatar identity, poster systems, and product mockups.

    line 58
  9. Distinctive:

    Add a concrete Definition of Done.

    line 59
  10. Distinctive:

    Do not default to a cinematic portrait, fashion editorial, or realistic headshot.

    line 69
  11. Very specific:

    If there is no reference image, prefer a 3 x 3 avatar exploration sheet first.

    line 70
  12. Very specific:

    If the user needs one final avatar, make it a platform avatar asset: large head silhouette, simple background, one memorable symbol, one …

    line 71
  13. Distinctive:

    Keep the symbol secondary to the face.

    line 72
  14. Distinctive:

    Prefer semi-realistic / semi-graphic / illustrated / badge-like treatments over photographic realism unless the user explicitly asks for a …

    line 73
  15. Distinctive:

    If a personal-brand avatar looks like a poster portrait, compare against examples/personal-brand-avatar-repair.md and repair toward …

    line 74
  16. Very specific:

    Default to a 4 x 4 expression pack.

    line 78
  17. Distinctive:

    Use "consistent imperfection": the same character can be wrong, but must be wrong in the same way.

    line 79
  18. Distinctive:

    Create a runnable eval pack:

    line 85
  19. Distinctive:

    Use it to test whether a generated prompt reconstructs expert-level capability from weak input.

    line 130

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

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.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.