Nate Herk | AI Automation says Reverse Engineer Anything helps nonprogrammers build more professional-feeling projects

For a motion-edit test, Claude Opus received a video by Jordan Watkins and a request to study it frame by frame, then make a skill for a similar edit. Herk says the original remained better in motion and sound design. With the reference to study, though, the agent produced a much better result than it could have without the example.

Herk also had Claude Code use the repository’s framework to analyze the repository itself, to understand how to use it and direct agents with it. The process starts by choosing a feature to investigate and considering how the app is built and what tools it uses. The agent reads a summary before moving into detail, then labels its findings as seen, inferred or unknown.

Those labels separate what the agent directly observed from what it inferred and what it could not establish. The process then moves into rebuilding, testing and iterating. Herk calls the framework practical and thoughtfully structured, and says an example gives an agent a clearer target than a general request.

The investigation of Loom generated more than a gigabyte of analysis before it was condensed into a shorter summary. Herk says asking an agent to reverse engineer an entire app is difficult; choosing specific features gives the investigation a focus. They also stress that some details may remain unknown when they cannot be seen in the available files or app.

Herk Recorder reproduces selected Loom features

Using the repository, the channel built Herk Recorder, a screen-recording app inspired by Loom. It has a camera bubble that can be moved and resized, plus a countdown before recording. The video shows a recording being stopped and played back.

Herk says the app reproduces selected Loom functionality rather than Loom’s full design. Recordings are saved locally in the app’s folder instead of Loom’s cloud. The example shows the framework being used to recreate particular features without reproducing the entire source app.

The channel also pointed the repository at Sand Castles to inspect its functionality. After connecting an API key, it made a version of its own; Herk says the result helps analyze Instagram content, including the channel’s own material.

For an AI evaluation studio called AIS eval studio, the channel examined how Langfuse displays and runs things. Herk says the channel then used some of Langfuse’s core features as a reference for its own version.

The repository-guided agent produced much better results

Herk compared two coding agents given the same prompt. The one without the repository finished first, while the agent using it took longer and produced what Herk described as much better results. Herk identified the investigation workflow as the difference between the two outcomes.

The repository had about 192,000 stars at one check. Roughly 30 minutes later, a refresh showed around 195,000 to 196,000. Herk links its growing popularity to the structured investigation process, while emphasizing that the underlying reverse-engineering idea was already familiar.

Herk also raised a possible cost advantage, conditionally. If Claude Sonnet 5.5 can use findings gathered during an investigation, they say, it might build something for a lot less than running Claude Opus 5.5 throughout.

The method extends beyond apps

Herk says results depend on the model doing the inspection and coding, and that different apps may require different setup. The repository does not guarantee an agent will build everything. They also warn against copying exact code or assets.

The review includes a segment promoting a free ScrollCraft website-design skill. Herk’s recommendation for this way of working goes beyond apps and software to building skills and improving day-to-day processes. They say the approach helps people without formal programming or software-engineering backgrounds make things that feel more professionally made.