デモ実行記録 · claude-sonnet-5-5 · 2026年10月

My image prompt keeps giving me messy, weird pictures. Here is the prompt: "beautiful stunning masterpiece 8k ultra detailed girl in a red dress standing in a forest at night, bright sunny day, holding a sword and a bouquet of flowers and a lantern, cyberpunk watercolor oil painting style, minimalist but very intricate, a sign that says 'WELCOME TO THE ENCHANTED WOODS', no bad hands" What I get: three different styles mixed, the girl has too many fingers, the sign text is gibberish, and it is never night and day at once. Please fix it. I'm not sure which tool I'm using.

Skill: image-prompt-debugging

Diagnosis

Problem	Evidence in your prompt	Confidence
Three mixed styles	"cyberpunk watercolor oil painting" names three media. The model blends them.	Likely
Never night and day at once	"at night" and "bright sunny day" contradict each other. The model picks one or averages them.	Likely
Messy and cluttered	"minimalist but very intricate" contradicts itself. Three held objects plus a sign plus a forest is a lot for one image.	Likely
Too many fingers	The girl holds a sword, a bouquet and a lantern, which needs several complex hand grips. Many generators handle hands badly.	Likely
Gibberish sign text	Many generators render long text unreliably, and a 5-word all-caps phrase is a hard case.	Likely
"no bad hands" does nothing	Many tools read "hands" as something to draw and ignore the "no". It can even draw more attention to hands.	Possible
Filler words add nothing	"beautiful stunning masterpiece 8k ultra detailed" is abstract and gives the model nothing to draw.	Possible

Minimal fix

Changes are in bold.

~~beautiful stunning masterpiece 8k ultra detailed~~ girl in a red dress standing in a forest at night, ~~bright sunny day,~~ holding a lantern only, ~~a sword and a bouquet of flowers and~~ watercolor painting style, ~~cyberpunk watercolor oil painting style, minimalist but very intricate,~~ a wooden sign reading "WELCOME" ~~that says 'WELCOME TO THE ENCHANTED WOODS', no bad hands~~

Clean rewrite

A girl in a red dress standing on a forest path at night, holding a glowing lantern in one hand, the other arm relaxed at her side. Full-body shot, centered, with tall dark trees behind her and a small wooden sign reading "WELCOME" to her left. Warm lantern light on her face and dress, deep blue moonlit shadows. Watercolor painting, soft washes, simple background.

If your tool has a negative prompt field, put this there: extra fingers, deformed hands, blurry text, mixed styles.

Variants to test

Night and day together. Keep the rest, but change the light to "dusk, last orange sunset glow on the horizon, deep blue sky above, lantern lit". Dusk is the closest a single image gets to night and day at once. Look for whether the light now looks coherent.
Keep the sword and flowers. Use "holding a lantern in her left hand, a sword sheathed at her belt", and drop the bouquet. Look for whether the hands stay correct when one hand is holding the lantern and the other is empty.

For a cyberpunk look, make that a separate test. Use "oil painting, neon pink and teal rim light" in place of watercolor. Don't combine it with the other styles.

Test plan

Change one thing at a time and generate 3–4 images per version with the same settings.
If hands are still wrong, crop or hide them: "hands behind her back", or a waist-up shot with the lantern at the edge of the frame.
For the sign, either shorten it to one word (as above) or leave it blank ("a blank wooden sign") and add the text in an image editor afterward. This is the most reliable method.

What I couldn't tell

Which tool you're using. This matters for several things:
- Midjourney and Stable Diffusion use negative prompts differently (--no and a negative field). Tools like DALL·E or ChatGPT image generation usually have no negative field, so describe what you want instead.
- Newer models handle text and hands much better than older ones, so your results may improve just by switching.
What your images look like. If you share one, I can tell you more precisely what went wrong.
Check your tool's docs for how it handles negative prompts, text rendering and aspect ratio. A tall ratio like 2:3 suits a full-body standing figure.

Real run in an isolated folder with only this skill installed. Only the Skill and Read tools were enabled, so no command was run and no file was written. The scenario is fictional. No image was generated: the model only wrote text prompts, so how well they work depends on the image tool you use. The model could not see any generated image, so its diagnosis is a set of hypotheses to test.

できること

デバッグのように進めます。正確なプロンプト、使っているツール、何が違うのかを集め、よくある問題の表と照らして読みます（一部が無視される、平凡な結果、ごちゃごちゃ、スタイルの混在、矛盾、構図の誤り、平板な光、色移り、崩れた手、文字化け、望まない要素、実行ごとの違い）。各問題について、プロンプト内の根拠と確からしさを示し、変更箇所を示した最小の修正、整理した書き直し、それぞれ 1 点だけ試す 2 つのバリエーションを示します。

ほかに扱うこと

テスト計画（変数を 1 つずつ変える、版ごとに複数枚を生成する、同じ設定で比較する）、ツールやモデル自体の限界である場合、画像やツールの挙動が見えないために判断できなかったこと。

向いている場面

「うまくいかない」プロンプト、いつも細部が抜ける画像、どの言葉が効いているかを知りたいとき。

補足とリスク

低リスク：スクリプトのない指示のみのパッケージで、ネットワーク接続やファイル書き込みはありません。プロンプトと結果の説明から判断します。画像を共有しない限り画像は見えず、ツールが言葉をどう重みづけするかも分からないため、診断はすべて試して確かめる仮説です。ツールの安全フィルターや規約の回避、実在の人物を欺く目的で真似ること、実在の人物や未成年者を含む性的な内容の作成には協力しません。ネガティブプロンプトや文字描画の挙動は、使用ツールの文書で確認してください。AIBars 制作のオリジナル（MIT）。架空の乱れたプロンプトで 1 回試用し、画像は生成していません。