Start with the missing evidence
Choosing Reverse Engineer Anything is not simply a question of target complexity. Ask what ordinary source inspection cannot establish. The official investigation skill directs REA toward questions requiring shipped binaries, packaged artifacts, decompilation or passive runtime evidence. For architecture questions about an ordinary repository with complete source, it explicitly recommends normal repository tools.
Choose a first step from your input
- Complete source: Start with definitions, imports, call sites and tests. Add artifact analysis when the claim depends on the actual shipped version.
- Packaged JavaScript: Identify the package format and readable content, then investigate one entry point or behavior instead of trying to recreate every module.
- Native binary: Identify the architecture and required engine, then investigate a function or data flow. Decompiled output still needs interpretation and validation.
- Only an interface demonstration: Record observable input and output first. A similar interface does not establish equivalent server algorithms, data models or complete behavior.
A skill installation is not tool readiness
The official skill distinguishes workflow instructions from callable tools. Installing instruction text does not register MCP or provide every engine. Confirm tool discovery, then check dependencies for the selected target. A task that only needs ordinary source analysis does not require an additional REA preparation sequence.
Additional cost should buy additional evidence
Use concrete success criteria: identifying an unknown entry point, explaining a reproducible failure, or grounding a claim in the shipped artifact. Tool count, output length and fluent prose are not sufficient measures of success. Establish a baseline with your current agent, then add REA under the same acceptance criteria and compare resolved questions and remaining unknowns.
MG’s video description reports a slower, more expensive single-screen run with REA, alongside an improvement judged by the author. This motivates evaluating cost and quality together, but remains one example we have not reproduced. A Reddit self-report also raises cost questions. Neither should be generalized to every program or model.
How to use our recorded case
Our frozen JavaScript case teaches input recording, structured output inspection and evidence checks. Its source is readable, so it does not establish that REA was necessary or superior to an ordinary coding agent. Use it to learn the recording workflow, then choose a task where artifact evidence is actually missing.
Use the comparison record template →
Sources and version boundary
- REA official investigation skill — Official documentation
- MG — How to Reverse Engineer Anything in Claude Code — YouTube description and chapters
- Reddit: REA discussion and a reported comparison — Community reports
The official link tracks a changing main branch. This site’s setup examples pin 6.5.0; check your installed version’s help and release notes for feature availability.
Continue your investigation
- Build a setup plan for your environment
- Inspect the reproducible case study
- Troubleshoot by symptom
- Browse all field guides