Source: raw/Reverse_Engineer_Anything_Does_Way_More_Than_You_Think.md

Creator: Austin Marchese (captioned “Austin Maresy”; spelling ^[inferred]) · URL: https://www.youtube.com/watch?v=lKz3cDrLgTk · Duration: not stated in the raw file · Platform: YouTube (auto-captions, fetched 2026-11-13). Sponsored by Clay; the creator also promotes his own tool, buildpartner.ai.

Austin Marchese takes the loop behind REA (“Reverse Engineer Anything”, the open-source tool that lets a coding agent inspect and rebuild software) and turns it into a five-step method for non-technical work: pick, gather, design, rebuild and grade, iterate. He applies it to four targets: a competitor’s funnel, a single feature of an app, your own repeated tasks, and an expert’s way of thinking. His point is that rebuilding a whole app (his demo cloned Apple Notes in about 40 minutes) is “genuinely useless”; the value is in copying the process, “not the property.”

Key Takeaways

  • REA is a harness, not new tech. He describes it as a wrapper that lets an agent drive existing tools (Ghidra, Hopper, IDA, Chrome DevTools) in a loop: give it a target, it inspects and traces, the agent builds, and it repeats until the app is recreated. Fuller coverage: Reverse-Engineering Software With Coding Agents.
  • The five steps: (1) pick what you are studying, (2) gather evidence, (3) design what you will build from it, (4) rebuild and grade it against the original, (5) iterate and make it yours.
  • Study small winners, not giants. “The bigger the player, the less you should study them”: Nike’s brand props up whatever it does, so pick “the small competitor that just broke out.”
  • Screenshots are the evidence pile. When Claude walks a competitor’s funnel in your browser, have it screenshot each key step so you see what it saw “instead of trusting its summary.”
  • Rebuild one feature, not the app. REA’s own README example targets one feature: understand how search works in the Notes app, show the evidence, build a similar feature. If the feature stands alone, build it as a Claude artifact rather than a hosted app.
  • Only rebuild with a reason. There “has to be a concrete enhancement” over the original; otherwise it is a “dopamine building loop.”
  • Mine your chat history for skills by asking what you had to send back. Ask Claude to review past conversations and flag skills to create. The tip: don’t ask “what did I use AI for?”, ask “what did I have to send back?” A task that came right first time probably does not need a skill.
  • Record and narrate a task to automate it. Record your screen in Claude while explaining what you do and why, “like you’re teaching a human on an onboarding call.” He says Claude then suggests programmatic routes (APIs, MCP connections) rather than replaying clicks. Next level: record yourself reviewing a finished output (what’s good, what’s wrong, why) as a rubric the skill can improve against.
  • Clone an expert’s method as a skill, then grade it. Gather their material (connectors for people you work with; public posts, videos and newsletters for people you don’t), build a Claude skill that captures how they decide, test it on a real problem, and compare it with what you know of their actual thinking. He frames the progression with the martial-arts model shu-ha-ri: obey, break, move away.
  • Limits. Neither REA nor a funnel walk-through sees the back end: customer support, email follow-ups, databases.
  • What stays hard to copy, in his view: the back end, relationships, trust, and the judgment about what is worth building. “If it’s Uber, it’s not the app. It’s the driver.”

Try It

  1. Funnel hack one rising competitor. Ask Claude to list your top competitors, pick a small one that recently broke out, and have Claude walk its funnel in your browser with a screenshot at every step. Then ask for actionable takeaways for your business, build what Claude can, and review the result.
  2. Run the “sent back” audit. Prompt: “Review my past conversations and list every task where I had to send your output back for corrections. For each, propose a Claude skill that would get it right the first time.” ^[inferred] (wording is ours; the “what did I have to send back” framing is his).
  3. Record one weekly task with narration, then ask Claude to propose an automation. Follow up with a second recording of you reviewing the output, and ask Claude to turn your comments into a rubric for the skill. Related first-party feature: Record a Skill (Cowork).
  4. Clone one feature, not an app. State the enhancement first (what the original can’t do for you). Gather screenshots or a screen recording of the feature, then ask Claude to build it as an artifact.
  5. Build an expert skill for a colleague or mentor from their reports, call transcripts and emails, and grade it: where does it get them right, what does it miss, what is it assuming?

Implementation

Tool/Service: REA (github.com/morluto/rea); Claude with browser access, screen recording and connectors; optionally buildpartner.ai (creator’s own product) and Clay (sponsor).

Setup: The video says to install REA “by running this prompt” but the prompt is on screen only. A search-result snippet of the REA README (2026-11-13, page not fetched) shows npx rea-agents@latest setup.

Cost: REA is open source. buildpartner.ai’s competitor research and expert-advice tools are described as free; Clay is “free to start” with a $250 credit offer for his viewers (sponsor claim).

Integration notes: Some sites block automated tools, so subpages of a funnel may be unreachable; browser walk-throughs are slow. Point REA and funnel analysis only at targets you are allowed to inspect (see the legal notes in the REA article).

Open Questions

  • The on-screen prompts (competitor identification, funnel scrape, takeaways, chat-history skill mining, social-media material gathering) are not in the transcript.
  • Screen recording “within Claude” is described without naming the surface. It may be Cowork’s Record a Skill ^[inferred]; not confirmed.
  • buildpartner.ai’s “expert advice” and social cloning tools are the creator’s own product; their quality is unverified.
  • Paul Graham quote (“There’s no better way to learn how something works than by trying to reproduce it”) is attributed by the speaker; the essay is not named.