“Clone this app” hides almost every decision that makes an investigation useful. Which behavior matters? Which package are we inspecting? What evidence would count as an answer? A smaller prompt makes those decisions visible.
The four details worth including
Give the agent a target, a behavior, the evidence you expect, and a stopping point. The stopping point is easy to forget. Without one, investigation can slide into implementation before you know whether the explanation holds up.
Use an exact local path where possible. Replace the bracketed text below; the examples are templates, not tool commands that can run unchanged.
Find the right place to start
Use REA to inspect [absolute local path].
I want to understand [one visible feature].
Identify the target format and the smallest relevant files or functions.
Show source locations or addresses for each candidate.
Do not launch the target or implement a replacement.
Finish by listing what you know and what you need to inspect next.This works well when your question is concrete but you do not yet know whether it belongs in browser code, an Electron bridge, or a native function. Let the artifact determine the route.
Trace an Electron feature
Follow [export/copy/search] through this extracted Electron package.
Find the renderer call, exposed preload API, IPC operation and receiver.
Match literal channels where possible and preserve dynamic unknowns.
Show the code locations supporting each hop.
Separate static relationships from observed runtime behavior.The useful output is a chain you can inspect. If the agent returns a diagram without source locations, ask for the missing evidence before moving on. Our first lab gives you a small chain to compare against.
Investigate a native calculation
In [binary path], investigate how [input] becomes [output].
Use my configured analysis provider.
Show the relevant function, its callers, instructions and constants.
If pseudocode disagrees with instructions, explain the discrepancy.
Do not claim recovered variable names are original source names.This prompt is especially useful for a numeric transformation, format check or small state update. It is less useful for “explain the whole application,” which can generate a great deal of output without answering a bounded question.
Ask for a testable reconstruction plan
Using only the established findings, describe the smallest implementation
that reproduces [specific behavior].
List input/output examples, boundary conditions and unresolved assumptions.
Propose tests that could disprove your explanation.
Keep observations, inferences and unknowns separate.
Wait before writing implementation code.The phrase “could disprove” changes the quality of the next step. A test that only repeats the happy path may confirm your preferred story while missing the condition that makes it wrong.
Stop an investigation from expanding forever
Pause broad exploration. Summarize the evidence collected for [question].
Name the single next probe most likely to resolve the remaining uncertainty.
State its expected result, alternative result, and stopping condition.
Do not inspect unrelated features.This is a scope tool, not a way around an agent’s rules. If the client refuses a request, clarify the legitimate target and purpose, or use a different permitted task. Do not treat a workaround prompt as evidence that an analysis is appropriate.
Keep a question log
Save the question, target version, relevant evidence references, open unknowns and next probe. The next session should start from that record rather than asking the agent to rediscover everything.
For a practical example, move through the Pocket Notes lab and try rewriting its conclusion with the evidence checklist.