I Tested the Viral Reverse Engineer Anything Skill. What You Need to Know.

Nate Herk | AI Automation14m 58sTranscript availableAdded Oct 14, 10:31 am GMT+8

Actionable Insights

Start with one observable behavior, such as dragging a recorder’s camera bubble, rather than asking for a Loom replacement. In a proposed feature-spec.md, record the input, expected output, failure cases, and features you explicitly exclude. Nate’s recorder example at 3:33 separates recording functionality from Loom’s appearance and cloud storage. Evaluate your version against those recorded behaviors, not against a screenshot alone.
Read the REA repository and choose an inspection path before setup. Its documentation distinguishes JavaScript/Electron inspection from native analysis using Hopper, Ghidra, or IDA. The documented setup command is npx rea-agents setup, from the repository rather than the video. Review the package and proposed configuration changes first, use a disposable project, and check the current runtime prerequisites. Installing investigation instructions alone does not install the complete tool connection.
Keep a proposed evidence.md with separate observed, inferred, and unknown findings. Attach file locations or test results to observed claims. At 9:39 the creator explains why treating inference as fact can corrupt the rebuild. Turn each inference that affects behavior into a test, and leave inaccessible server behavior unknown rather than asking the model to invent an implementation history.
Save the investigation summary and retrieve detailed artifacts only when a specific question requires them. At 8:38–9:39, the creator describes a Loom analysis exceeding a gigabyte and warns against repeating it. That size is his report, not a measurement from this run. Record the target version and artifact locations so a later agent can reuse evidence without assuming an old analysis describes a changed app.
Build a comparison suite before declaring parity. For the recorder example, proposed checks include pause/resume, bubble movement and resizing, window selection, and playback of the saved local file. Record expected and actual results. Claude Code’s verification guidance supports tests and screenshot comparisons, but passing these checks would establish only the tested behavior, not equivalence to all of Loom.
Check authorization and data exposure before inspecting a commercial app. Prefer your own software or an explicitly permitted target for a first experiment. REA’s local analysis does not imply that an agent’s model provider receives no tool output. Avoid private customer data, copied assets, and access-control bypasses. For redistribution or disputed license terms, obtain advice for the relevant jurisdiction rather than treating local possession as permission.

Core thesis

A reference product gives a coding agent a more specific target than a generic request. REA adds a structured investigation process and real inspection tools. The useful outcome is an evidence-backed reconstruction of selected behavior, not a promise that an agent can recover every product’s implementation.

The video is strongest when it narrows scope and distinguishes seen, inferred, and unknown evidence. Its broad framing of reference-based design as decompilation blurs two different activities. Studying a finished animation is not the same technical process as translating compiled machine instructions into pseudocode.

Key ideas and timestamped moments

1:00–2:32. Nate compares human-edited reels and generated versions, then acknowledges that the reference motion edit still has better motion and sound design. This supports using examples to communicate intent, not the claim that AI cannot produce original work.
3:03–4:34. Herk Recorder, a content-analysis app, and AIS eval studio illustrate selective reconstruction. In his account of the demo, the recorder stores files in a local folder, whereas he describes Loom’s cloud storage. That architectural difference matters more than visual resemblance.
7:36–8:38. Pick a feature and route the investigation according to the application’s technology. This turns an unbounded cloning request into an inspection task with a stopping condition.
8:38–10:40. Reuse summaries, separate evidence categories, and test the rebuild. These are the most transferable parts of the workflow.
11:10–12:41. He warns against copying exact code or assets and notes that results depend on the coding model. His model-name examples are hypothetical workflow discussion, not verified product specifications here.
12:41–13:42. He reports that REA-assisted agents took longer but produced better results. There is no disclosed controlled benchmark in the extracted transcript, so the result remains his experience.

Comment insights

The extracted comments contain a useful disagreement about what is new. @andycted-to-dance, with 7 likes, sees packaged prompting and planning. @Zuranthus, also with 7 likes, emphasizes real decompilers and repeated comparison. The repository supports the latter’s distinction between tool-backed inspection and guessing, but not a universal guarantee of exact reconstruction given enough time.

@dmon1088 adds a maintenance objection that the main demo underplays. A weekend clone may lack support and accountability even when its interface works. The same commenter offers color-grading math as a use case where recovering precise behavior could matter more than visual imitation. That is a proposed application, not a demonstrated result.

@SolutionsJoe warns about EULAs. This usefully challenges the video’s casual association between possession and public availability, but the comment’s broad legal conclusion is not established for every contract or jurisdiction. @brokenhero6828’s question about features versus full applications shows why the scope distinction should come before installation advice. Off-topic promotion and personal insults add no technical evidence.

Deep research

Sources were retrieved on 2026-10-14. Web search was unavailable, but direct web retrieval worked.

REA is more than a prompt collection

The official repository describes an MCP connection plus matching workflow instructions. It documents native pseudocode and assembly, JavaScript module relationships, website observations, and other target-specific outputs. Native analysis can use Hopper, Ghidra, or IDA; static JavaScript inspection does not require a native engine. This supports the video’s technology-routing advice and corrects any impression that downloading a skill alone supplies all capabilities.

The same source distinguishes static inspection from runtime capture, which interacts with targets using the user’s permissions. It also states that analysis is local while the agent receives results. These are meaningful operating limits. They contradict an interpretation of REA as either an unrestricted remote-code recovery service or a tool that necessarily keeps every finding away from model providers.

The README cites a DX-Ball reconstruction passing 3,205 original-x86 cases and reproducing all 63 compiled function bytes. This is a project-reported, narrow example, not an independently reproduced benchmark here. It supports the possibility of precise reconstruction under suitable conditions, not the commenter’s claim that any feature eventually becomes exact.

Examples and tests help, but do not settle creativity or speed

Anthropic’s best-practices guide recommends specific context, reference patterns, screenshots, and checks an agent can execute. It also recommends exploration before implementation when a task is uncertain, while warning that planning adds unnecessary overhead for small, clear changes. This supports Nate’s practical method without establishing that every task benefits from a full REA investigation.

Neither that guide nor the video supplies a controlled test of the claim that AI cannot be creative. The creator’s own admission that the reference motion edit remains better is a more defensible, example-specific conclusion. Likewise, his faster-versus-better comparison lacks task counts, scoring criteria, model controls, and cost records.

Local possession does not resolve legality

17 U.S.C. § 1201 prohibits certain circumvention of access controls and includes a conditional interoperability exception in subsection f. That exception has requirements, including lawful access and a specific interoperability purpose. It is not general permission to clone products. This U.S. source complicates the video’s reassurance. It also shows why the comment’s claim that reverse engineering paid software is usually illegal requires more analysis than citing a EULA alone. It does not determine the applicable law or enforceability of a particular EULA.

Verdicts

Agree, high confidence, that feature-level scope and explicit evidence categories improve the clarity of an investigation. The transcript and REA’s documented outputs support the mechanism. The practical payoff is a testable specification, not automatic parity.
Mixed, high confidence, on the claim that REA is nothing new. Reference-based reconstruction is the video’s existing practice, but the current repository supplies actual inspection integrations beyond a prose prompt. Treat the package as tooling plus procedure.
Mixed, medium confidence, on better results at greater time cost. Nate reports that pattern, and deeper inspection plausibly explains it. There is no controlled evidence here for its frequency or magnitude. Benchmark your own feature with equal budgets and the same acceptance tests.
Disagree, high confidence, with a universal reading of AI’s inability to create. The evidence only compares selected outputs. It cannot establish an impossibility claim.
Mixed, high confidence, on the legal reassurance. Avoiding stolen code and assets is sensible, but having a local copy does not settle access-control, contract, or redistribution questions. Obtain target-specific permission and advice.

Screen-level insights

Two extracted frames were visually inspected. At 0:00, an X post displays a REA introduction with an MCP description, a repository-of-the-day badge, and setup material. This establishes the promotional context for the opening, not a live technical test.

At 3:33, the screen shows an “Original vs ours” comparison with the screen-recorder tab selected. Loom appears on the left and Herk Recorder on the right, with visibly different interfaces. The nearby narration discusses moving and resizing the camera bubble. The frame supports the distinction between appearance and selected functionality, but a still image cannot prove pause/resume behavior or saved-file correctness.

No later extracted frames were visually audited in this run. The detailed evidence-tagging and testing discussion above comes from the transcript, not an invented view of code on screen.

Verification notes

After the unslop editorial pass, the synthesizing agent completed four separate role-based reviews: source/evidence, transcript/comment/frame fidelity, hallucination/overclaim, and Actionable Insights. These were not independent agent reviews. The reviews checked source attribution, technical details, quotations, numbers, confidence, uncertainty, and the distinction between observed results and proposed experiments. The Actionable Insights audit checked first steps, links, evaluation criteria, and integration cautions. Corrections clarified local-folder storage and preserved the legal commenter’s qualified claim rather than turning it into an absolute. Remaining limits include a disabled search provider, no executed REA benchmark, no independent confirmation of the creator’s performance comparisons, and visual inspection limited to the two frames described above. Official documentation establishes advertised mechanisms, not measured success on this video’s projects.