You've seen the screenshot: a Photoshop effect you'd like to recreate somewhere else, and an AI coding agent that's happy to try. Whether it can do anything useful depends less on the model than on what it knows before it starts. That's the whole idea behind Reverse Engineer Anything, a small site I run with one guide per application.
REA itself is an open-source toolkit that helps AI coding agents investigate how software works. My site is an independent set of guides for it, separate from the toolkit and from Adobe and Autodesk. It doesn't run REA and doesn't accept uploads. The part worth sharing is the checklist each guide pushes the agent through.
1. Is it an export or the editable file?
On the Photoshop guide, the page separates two kinds of input. PNG, JPEG and TIFF show appearance. A PSD or PSB may expose editable structure such as layers or smart objects. A flattened image can't tell you whether editable layers, smart objects or plugins were involved at all, so if all you hand over is a picture, the honest answer is a plan with unknowns, not a recipe.
2. Which exact environment?
The same guide asks the agent to record the Photoshop release, whether it's Windows, macOS or the web form, colour mode, bit depth, plugins and cloud features, and not to merge desktop and web behaviour. It sounds fussy until you remember that "the blur looks different" can come from any of those.
The AutoCAD guide is the same idea for drawings: version, units, layers, blocks, dimensions and external references. It says to compare outputs on a copy rather than judging by appearance.
3. Lawful input, copies only, and a stop condition
The guides I've linked here tell the agent to use only files and access you're allowed to use, to keep originals and work on copies, and to stop when the target or input can't be inspected. It also asks for a report of what was actually observed versus what is still unknown. The guides don't include the software, an installer, vendor code or a license, so you need your own authorized copy when a task calls for one.
What you actually download
A guide is a package you hand to your own agent: a SKILL.md, REA setup instructions and a target reference for that application. New accounts start with 1 credit valid for 7 days, each premium guide costs 1 credit, and repeat downloads are free. That's all the product is: context, written down once, for an agent that otherwise starts from nothing. It won't make a task possible that the files and tools can't support, and the guide says so itself: it's a focused starting point for an investigation, not a promise that every version or goal will work.
Disclosure: it's my site.
If you've asked an agent to reproduce behaviour from a finished export, what context did you wish you'd given it first?
Top comments (0)