Tested ChatGPT Images 2.0 this weekend. Here's what I learned.

First attempts: boring. Generic AI illustrations, nothing close to the examples on OpenAI's website.

Then I rewrote (well, Claude rewrote) the prompts with proper detail.
Specific composition.
Specific lighting.
Specific layout.
Specific text rendered verbatim.
The results changed.

So, the model is good, but I wouldn't say perfect. Character consistency is still not there. But text creation is indeed really good.

If you want better outputs, definitely take into account the recommendations from OpenAI's own developer guide:

🖼️ Write prompts in a consistent order: background → subject → details → constraints. Treat it like a creative brief, not a search query.

🖼️ For photorealism, use the word "photorealistic" directly. Add "real photograph", "natural lighting", "shot on film". Avoid words that imply studio polish unless you want studio polish.

🖼️ Put any text in the image inside quotation marks. Specify the font style, size, and placement. Spell tricky words letter-by-letter if needed.

🖼️ Be specific about composition. Close-up or wide. Eye-level or low-angle. Soft diffuse or golden hour. The model needs to know what shot you want.

🖼️ State what should NOT change. "No watermark." "No extra text." "Preserve identity, geometry, layout." Especially important for edits.

🖼️ Iterate instead of overloading. Start with a clean base prompt. Refine with small single-change follow-ups. Easier to debug, better results.

And the skill isn't using AI image tools.
The skill is knowing what you actually want to see.

Have you tested it?