[00:00]
Deepseek plus Rea can now take apart almost anything on GitHub and [music] rebuild it as a tool for your business. Rea stands for reverse engineer anything. I plugged it into Deep Seek Harness [music] and pointed it at three
[00:14]
famous open-source projects, an agent library from Hugging [music] Face, a page builder with 13,000 stars, and the forum software behind thousands [music] of online communities. Deepseek opened each one up, showed me the evidence of
[00:28]
how it works, and then built me something new from it. A code-based agent that writes client reports, a launch page builder, and a radar that finds people asking for help inside a big group. Every build in this video
[00:41]
came from a question someone asked in the AI Profit Boardroom this week. I'll show you each one working, the exact prompts, and the setup step by step. So, what is Rea? It's a free open-source tool on GitHub, MIT licensed with over
[00:56]
467,000 stars. Version 6.9 came out on the 12th of November. It gives your AI agent a set of reverse engineering tools. Native programs, JavaScript and Electron desktop apps,Net apps, and websites. Its
[01:12]
static tools read JavaScript and TypeScript directly. And when a project is in another language, DeepSeek reads the code itself and REA still keeps the evidence trail. Everything runs on your own machine. And here's the part I love.
[01:27]
Every answer comes with the evidence behind it and a list of what it could not prove. Now, the brain DeepSeek Harness is Deepseek's own agent app. It's open source, too, with over 280,000 stars. Everything inside it is a plugin,
[01:42]
so REA just slots straight in. I'm running DeepSseek V4.1 Flash inside it. It's fast and it can read screenshots which helps when it checks its own work. Here's how it works in practice. You open Harness, you pick a workspace
[01:57]
folder and you drop the GitHub project you want to study into a folder called source. Then you write one prompt in four parts. Study, evidence, rebuild, polish. Study tells REA what to take apart. Evidence tells it to write down
[02:13]
what it found. Rebuild tells DeepSeek what to make from it, and Polish tells it how it should look. Why does this matter? Because the best ideas in software are already sitting on GitHub for free. Someone spent years working
[02:27]
out how a good agent loop works or how a page builder stores a page. With Rea, your AI reads all of that for you, explains it with proof, and turns it into something you can actually use in your business. Quick one. If you want my
[02:41]
best setups for DeepSeek and REA, like the three prompts I'm about to show you, the evidence checklist, and the glass lab design brief that made these builds look this good, they're all inside the AI profit boardroom. That's also where
[02:54]
all three of these questions came from. You get a community of people building with AI every day and a new tutorial almost every day. The link is in the comments and the description. Build one, the client report agent. A member asked
[03:08]
this week, "Are there any lessons on building codebased agents, not NATN?" So, I gave DeepSeek small agents. It's hugging faces agent library with over 30,000 stars, and its whole idea is agents that think in code. Here's my
[03:23]
prompt. Use Rea to reverse engineer how a small agent's code agent works, how it plans, writes a small Python step, runs it, reads the result, and decides the next step. write the evidence to evidence.mmd, then build my own version,
[03:39]
a client report agent for a marketing agency. Watch it go. It loads the reverse engineering skill, maps the project, and walks the agent loop file by file. And here's something I really like. Small agents is written in Python,
[03:53]
and REA's static tools read JavaScript. So, Rea scanned the project, found zero JavaScript files, and said so straight up, no guessing. So, DeepSeek read the Python itself and pinned every claim to a file and a line number, the loop
[04:08]
driver, the planner, the step that writes the code, the executor that runs it, and the tool that ends the run with a final answer, seven key files, and a section called what Rea could not prove. Then, it builds and here's the agent
[04:24]
running. Pick a client, hit run agent, and each step appears as a card. the thought, the little bit of code it ran, and the result. Read the data, compare the weeks, find the wins, write the report. One honest note, Deepseek
[04:39]
invented a local clinic as one of the demo clients, so I swapped that name for a coaching business before filming. The agent and the layout are exactly what it built. At the end, the finished client report slides in. Headline wins, three
[04:53]
charts that draw themselves, and next week's actions. That's the same loop a real code agent uses rebuilt for an agency. You could point it at your own client data and Monday's reports write themselves. Build two, the launch page
[05:08]
builder. Another member asked, "How does Julian make the websites for his launch kits?" So, let's take apart a page builder. Puck is an open-source visual editor MIT licensed with over 13,000 stars. My prompt. Use REA to reverse
[05:25]
engineer how Puck builds pages. How blocks are registered, how the page is stored as data, how drag and drop changes that data, and how the page renders from it. Write the evidence, then build my own launch page builder
[05:39]
for course and community owners. This time, Rea could go allin because Puck is written in Typescript. Deepseek bundled Puck's core into one JavaScript file and Rea tore through it. 32,000 pieces of code, 81 findings, over 1,500 links
[05:58]
showing which function calls which. And it found the two functions that matter. Insert action, which drops a new block into the page data, and move component, which handles drag and drop. A block goes in, the page data changes, and the
[06:13]
page draws itself again from that data. And here it is on the left, a library of blocks. Hero, promise, what's inside, bonuses, testimonials, questions, and a
[06:25]
join button. Click a block and it flies onto the page. Or hit build my launch page, and the whole thing assembles itself block by block. On the side, you can see the page data updating live. That's the exact pattern Puck uses.
[06:40]
Blocks then data, then the page. If you run a course, a community, or a coaching program, this is how you get a launch page up in minutes instead of a weekend. Build three, the question radar. This one came from a member who runs a huge
[06:55]
online group, close to 200,000 people. They asked, "Can AI help me find the people in my group who actually need help?" So, I gave DeepSeek part of Discourse. That's the open- source forum software behind thousands of communities
[07:10]
with over 48,000 stars. I gave it the search code and the solved plugin, the part that marks a question as answered. My prompt, use to reverse engineer how
[07:22]
discourse finds and ranks topics, how search scores posts, how it tells unanswered topics from solved ones, and which signals feed in. Then build a question radar for a business coach. It started by listing exactly what was in
[07:37]
the folder. So the evidence begins from what's really there. Then it went through the search class, more than 1,700 lines, and found how discourse ranks results by relevance. And inside the solved plug-in, it found one
[07:51]
function that sorts every topic into three buckets: resolved, in progress, or unanswered. That's the signal a business owner actually wants. Who asked a question, and who's still waiting for help? Here's the radar. 42 posts from a
[08:06]
big group. Each one a dot on a glass dish. Hit scan the group and the beam sweeps round. The posts from people who need help rise up and glow. Click one and you see the post. Why it scored high and a short reply draft. Helpful advice
[08:22]
first. No hard selling, which is exactly how you should show up in a group anyway. For a coach, a consultant or an agency. That's your best conversations found for you every morning. Now, another member asked a great question
[08:36]
this week. What evidence should an agent have before it says the work is done? This is where REA shines. Every run ends with an evidence file, the files it read, the functions it found, the order things happen in, and then a section
[08:51]
called what it could not prove. That last part matters most. If an agent tells you what it doesn't know, you can trust the rest. So, here's my checklist before I trust any build. One, open evidence.md and check it names real
[09:06]
files and functions. Two, read the limits section. If something important is in there, ask Deepsee to dig into that part next. Three, open the build and click every button yourself. Four, check the build uses the logic it found,
[09:21]
not a fake version that just looks right. Five, keep the source folder read only so your agent studies the project and never edits it. Let's set it up. Open your terminal and run npxa agent setup-client
[09:37]
dsh that registers inside deepsee harness and installs the reverse engineering skill. Rea needs node 24. If setup stops and asks for node, install it and run setup again. Then run npx rear agents doctor. You want the
[09:54]
deepseeek harness line marked healthy. Now start harness with DSH web. Add a workspace folder. Make a folder inside it called source and put the GitHub project you want to study in there. A shallow clone is plenty. Open a new
[10:08]
session in that workspace. Check the model says Deepseek V4.1 Flash and paste your four-part prompt. Study evidence rebuild polish. Then let it work. Mine
[10:20]
took a few minutes per build. When it's done, open evidence.md first, then open index.html in your browser. Here's which one to build first. If you run an agency, start with the client report agent. If you sell a course or run a
[10:35]
consultant, start with the question radar. And a few mistakes to avoid. Don't point it at a whole giant repo on your first go. Give it the part that
[10:47]
does the thing you care about, like I did with Discourse. Always check the license so you know what you can reuse. And rebuild the idea in your own words and code. Never copy someone's product and call it yours. That's DeepSeek Plus
[11:01]
Rea. Three famous GitHub projects taken apart with evidence and rebuilt into tools for agencies, course owners, and coaches. If you want the exact prompts, the evidence checklist, the glass lab design brief, and the full setup guide,
[11:15]
they're inside the AI Profit Boardroom. The link is in the comments and the description. See you inside.