Once it is installed, the first sentence you type decides how the rest of the session goes.

The question: one feature, plain language, a checkable output

The phrasing the project uses as its example: understand how search works in the Notes app, show me the evidence, and build a similar feature for my project.

Four elements are doing work there. A specific target, not a category. A single feature, not "the whole app." An explicit demand for evidence, which rules out a confident story. And a deliverable, so the investigation lands somewhere.

The anti-pattern is "reverse engineer this software for me." The scope is so wide that the best available answer is a feature inventory, and you are still the one guessing where to start.

The evidence chain: work downward, do not ask for everything

REA works conversationally, and the right rhythm is several rounds of narrowing.

Round one is the overview: which modules exist, where the entry points are, which files relate to the target feature. Round two follows one thread: from the button in the interface to its handler, then to whatever that handler calls. Round three asks for detail: the pseudocode of one function, the purpose of one constant, the origin of one string.

Every finding carries a file offset, a confidence level and the unknowns the tool could not resolve. When it reports an unknown, do not skip past it. That is usually the most valuable place to push. If it says "this function handles percentages, but the exact operator dispatch was not confirmed," the next question is the dispatch.

One signal of real quality: does it volunteer counterexamples. If it claims a branch is only taken for multiply and divide, ask what addition does instead. That is exactly how the Calculator case resolved, where adding 10 percent and multiplying by 10 percent produce different numbers from the same keystrokes.

The cost: 207 tools cut both ways

REA exposes 207 tools, roughly 179,000 tokens by cl100k estimate. That will eat most of a context budget.

Three ways to keep it in bounds. Use the default compact input schema, and only switch to full when you need the field guidance. Constrain the engine to what you actually have, because routing to a provider that is not installed burns turns. And do one feature per session rather than three in parallel.

Accept one fact of life: a full investigation issues dozens of tool calls. The independent test called this out as a real burden. So start cheap. Static JavaScript analysis needs no native engine and returns structure in under a minute; see whether it answers your question before reaching for Ghidra.

Three red lines: contract, circumvention, jurisdiction

Doing something technically is not the same as being allowed to.

The first line is contract. Many commercial apps prohibit reverse engineering in their terms of service. That is a breach of contract rather than copyright infringement, but it is a real exposure, and it does not go away because your purpose seems reasonable.

The second is anti-circumvention. Section 1201 of the DMCA prohibits getting around access controls, the DRM and encryption protecting the binary itself, even where reverse engineering the underlying software would otherwise be permitted. REA provides no circumvention, but if you wire in other tools to do it, the risk is yours.

The third is jurisdiction. The same act is treated differently in the United States, the European Union and the United Kingdom. Reverse engineering for interoperability has consistent case support in several countries, but the boundaries differ, and India and other jurisdictions protect code as a literary work on a different logic again.

REA's own disclaimer draws the line tightly: lawful research, education, interoperability, software preservation, and software you own or are authorised to examine. Cracked commercial tools, stolen licence keys and bypassing licence restrictions are explicitly outside it.

Worth adding: Anthropic has started restricting Claude for game modding and decompilation. Switching agents does not fix a compliance problem, it only relocates it.

Do not mistake rebuilding a feature for rebuilding a product

This matters more than the law in everyday practice. What you recover is usually a client-side behaviour. The real logic often lives on a server; with something like Gmail you cannot see the code at all and can only black-box it. Even a faithful feature rebuild leaves updates, integrations, customer data and support, which is most of what a product actually is.

The sane expectation: understand how one feature was built, then build your own version your own way. Not carry someone else's product home.

One note on division of labour

If the goal is to understand, use REA. It reads code and shows evidence. If the goal is to build your own version of an app, that is a different pipeline, closer to the four-stage flow in craft-anything, which by default refuses to read source and works only from what a user can see. Do not mix the two jobs in one session.