In this week's GitHub roundup, Andrew Warner and Adam move from reverse-engineering apps to making videos, diagrams, CAD parts, and everyday software with agents. The common thread is not “AI can do everything.” It is that a specialized tool can give an agent a better workflow and more inspectable output than a broad prompt alone. The ten repos below are grouped by use, rather than episode order; the chapter list follows the video. Supporting videos and X posts from the creator's companion page appear with each repo, with direct links if an embed is blocked. Billion Context has no additional media on that page.
Start with the job, not the star count. REA and Ghidra help inspect software you are authorized to analyze; HyperFrames and Diagram Design help produce editable visual assets; Text to CAD gets you a prototype you can physically test. None removes the need to verify behavior, rights, or quality.
Watch the Full Episode
Credit: Andrew Warner and Adam present the projects and demonstrations in the original episode. Their October 16 companion page supplies more videos, X posts, and creator links. The episode is presented by Zapier; Zapier MCP is the sponsor reference, not one of the ten repos.
1-3: Inspect and Rebuild Software
1. REA: agent-guided inspection
Repository · Guides · 00:21 demo
REA connects a coding agent to inspection workflows for supported binaries, apps, websites, and runtime behavior. Its own README says findings should include evidence and limitations. Andrew shows an agent inspect a 2048 game and build a small working copy (03:27). That is a feature reconstruction, not proof that it can faithfully copy any full app. Adam also notes that the project lowers the friction of techniques that already existed. For a countercheck, the episode cites a Mac Preview comparison where REA did not clearly beat a plain agent. The creator's REA X discussion illustrates the hype; do not use a social post's star count as a capability test.
Boundary: inspect only targets you own or are authorized to analyze. A software license, platform terms, third-party rights, and local law may restrict reverse engineering or redistribution. Do not use an agent to pull secrets or bypass access controls.
Watch the demos and view the discussion
2. AnyPS5: executable relinking, not a universal game port
Repository · 05:11 discussion · Related X post
AnyPS5 aims to convert PS5 executables into native programs rather than emulate an entire console. The hosts are careful: it is a fast-moving project, and encrypted game data and compatibility still block the “run any game tonight” interpretation. Its repo is the place to check current supported inputs and limitations. The companion page also links an end-to-end demonstration using a free, open-source test game rather than a pirated commercial dump.
Watch the demos and view the discussion
3. Ghidra: evidence from a compiled program
Ghidra is the NSA's open-source software reverse-engineering framework. Andrew's controlled demo builds a tiny vault app with a hard-coded code, decompiles it, and finds that code in the compiled program. The lesson is about why a secret embedded in client software is not a secure secret. It is not a demonstration of breaking a properly designed password system. REA can use Ghidra as one of its analysis engines; Ghidra itself remains the lower-level tool (16:07).
Watch the demos and view the discussion
4-5: Produce Visual Work with More Control
4. HyperFrames: render HTML as video
Repository · Documentation · 10:10 demo · Related X post
HyperFrames lets an agent author an HTML/CSS composition, preview it, then render frames and audio to a video file. Andrew's team makes a short vertical promo, while Adam argues that the separate framework gives an experienced creator more control over motion, timing, and revision than a one-click video prompt (11:41). The project docs describe the CLI, reusable components, and local render path. “Open source” describes the framework, not a promise that generation or render hardware costs disappear.
Watch the demos and view the discussion
5. Diagram Design: brand-aware diagrams
Repository · Project site · 12:46 demo · Related X post
Cathryn Lavery's agent skill creates self-contained HTML/SVG diagrams and can learn a site's colors and fonts. The value is a constrained visual vocabulary that is easier to edit and keep on-brand than an unconstrained image prompt. The host shows architecture, flow, and other diagram examples. The README and gallery are better starting points than a fixed number of templates: the repository's description and body currently disagree on the count, and that catalogue changes.
Watch the demos and view the discussion
6-7: Keep Context and Make Physical Parts
6. Billion Context: a compaction layer
This project tries to keep coding sessions useful for longer by compressing prior context. Adam's objection at 18:15 is the important test: a summary can retain the broad plan and lose a critical detail. A “billion-token” name is not a billion-token active context window. Keep source files and decision records retrievable, and ask the agent for exact evidence when a detail matters. The repository's token-savings and duration claims are project claims, not results independently reproduced in this episode.
The companion page does not include a separate video or X post for Billion Context. The linked episode segment is the relevant demonstration.
7. Text to CAD: prototype, print, measure, revise
Repository · Project site · 20:31 demo · Related X post
The tool gives coding agents local workflows for CAD and exports such as STEP and STL. Andrew asks for a phone stand with a slanted back and front lip, and the agent produces an STL and preview. Adam's strongest point is the iteration loop: print a draft, measure the mismatch, photograph it, and revise (21:46). A plausible render does not prove fit, printability, strength, or safety. Use the repo's design-for-manufacture checks as aids, then inspect and test the actual part before making or selling multiples.
Watch the demos and view the discussion
8-10: Build and Own Better Apps
8. Matt Pocock's Skills: make the process explicit
Repository · 06:29 TDD segment · Related X post
Matt Pocock's small skills cover interviewing a user, test-driven development, debugging, and review. Andrew demonstrates a bill splitter, but Adam catches a missing visual step: genuine TDD should show the failing tests before implementation, then the changes that make them pass (09:32). A passing final suite alone does not prove the process or the specification was right. Treat community skills as executable instructions: read them, scope access, and keep the human accountable for requirements.
Watch the demos and view the discussion
9. SwiftUI Agent Skill: current platform guidance
Repository · 23:09 segment · Creator walkthrough
Paul Hudson's skill brings SwiftUI API, accessibility, performance, and deprecation guidance into an agent workflow. It is a useful counterweight to stale model memory, not a replacement for the current SDK, Xcode compiler, device tests, or accessibility review. The episode only briefly introduces it; the README and author's walkthrough give the fuller setup.
Watch the demos and view the discussion
10. openGym: a self-hosted workout tracker
Repository · 24:18 segment · Related X post
openGym is a self-hosted tracker for routines, workouts, and progress, with imports from popular fitness apps. Its appeal is ownership of the app and data, not a claim that every paid fitness service can be replaced without effort. The README documents a Docker Compose setup and an important license distinction: exercise images and animations have separate use terms from the app code. Review those terms before repackaging or commercial distribution.
Watch the demos and view the discussion
A Sensible First Test
- Pick one authorized input: your own app binary, a public example repo, an owned thumbnail, a synthetic diagram brief, or a low-risk printable part.
- Write acceptance criteria first: what must be correct, editable, reproducible, and licensed for your intended use? Include one failure case.
- Read the repo before installing: check its README, license, dependency setup, external calls, permissions, and model costs. Isolate untrusted binaries and third-party agent skills.
- Run one small job: record setup time, human corrections, output quality, and whether the result still works on a second attempt. Stop if the tool creates more review or risk than it saves.
Video Chapters
| Time | Segment | Time | Segment |
|---|---|---|---|
| 00:00 | Top repos | 00:21 | REA |
| 03:27 | REA 2048 demo | 05:11 | AnyPS5 |
| 06:29 | Matt Pocock's skills | 09:32 | Why skills help |
| 09:42 | Zapier sponsor | 10:10 | HyperFrames |
| 11:41 | Video demo | 12:46 | Diagram Design |
| 14:40 | Ghidra | 16:07 | REA with Ghidra |
| 17:00 | Billion Context | 18:15 | Compression limits |
| 20:31 | Text to CAD | 21:46 | 3D printing loop |
| 23:09 | SwiftUI skill | 24:18 | openGym |
| 25:04 | Why own an app? | 26:04 | Closing goal |
Business Idea Prompt: Prove One Repo's Value
A small, paid-worthy workflow needs a buyer and a repeatable outcome, not just an impressive repo demo.
Build, test, and interview
One buyer, one job, one honest result.
Source Map and Credits
Episode and editorial context: Andrew and Adam's video, supplied transcript, and their companion page. Project behavior and limitations: the ten linked GitHub repositories, especially REA, HyperFrames, Text to CAD, and openGym's media-license note. The supporting video and X embeds link to their creators and are examples, not independent proof of general performance. Repo features, licenses, install steps, and star counts change; evidence checked October 16, 2026.