Prompt coach for Photoshop generative fill

A photographer or designer runs generative fill, gets a weird result — a floating arm, a sky that doesn't match the lighting, a banner that says 'Hapyy Birthday' — and has no idea whether the problem is their prompt, the model's limitations, or just bad luck. They re-run it five times and get five different wrong answers. There's no feedback loop.

Adobe has no structural incentive to fix this: their job is to ship the generation model, not to teach users how to get good results from it. The gap between 'what users type' and 'what the model actually responds to' is a prompt-engineering problem, but Adobe's UI treats prompting as self-evident. It isn't. The complaints about needing 'a lot of extra direction' and generations that 'get confused easily' aren't model failures — they're prompt failures that users can't diagnose.

What's missing is a layer that interprets what a user is trying to do — 'remove the person from the background and fill it with matching grass' — and translates that into the specific phrasing, structure, and constraint language the underlying model actually responds to. This isn't a chatbot. It's a structured prompt builder that understands generative fill's specific failure modes: the text distortion problem, the random-object-replacement problem, the lighting-mismatch problem. It surfaces those as guardrails before the user hits generate, not after they've wasted ten credits.

Without this, users either burn through generation credits on retries or give up and manually retouch — which defeats the point. The cost is real: Adobe charges for generative credits at volume, so bad prompts are literally money wasted. And the problem recurs every single session because prompting is never the same twice.

What to build

Build a browser extension and Photoshop plugin that intercepts a user's generative fill prompt, parses the intent, flags known failure patterns (text generation, object replacement, edge lighting), and rewrites or structures the prompt before submission — with a visual 'confidence score' based on what the model handles well vs. poorly.

Where to start

Start with the single most-complained-about failure: AI-generated text on banners and signage, where the misspelling problem is consistent and predictable enough that a deterministic fix (detect text intent, force character-by-character constraints) wins immediately against the raw model.

The hard part

Adobe could add prompt guidance natively with a single sprint, and if they do, the wedge disappears — the hard part is building fast enough and making the prompt logic specific enough that it's genuinely better than whatever Adobe ships, before they ship it.

How it makes money

Monthly subscription per seat, priced below one wasted hour of a freelancer's time — around $9–15/month — with a free tier limited to text-generation prompts to drive word of mouth among event and commercial photographers.