ConvertKit AI

ConvertKit AI AI

ConvertKit AI AI

ConvertKit AI AI

ConvertKit AI AI

ConvertKit AI AI

ConvertKit AI AI

ConvertKit AI AI

ConvertKit AI AI

ConvertKit AI AI

ConvertKit AI AI

Claude Code built its own software for a little smart car I'm building.

TLDR: Check out the video # Box to Bot: Building a WiFi-Controlled Robot With Claude Code in One Evening I’m a dentist. A nerdy dentist, but a dentist. I’ve never built a robot before. But on Sunday afternoon, I opened a box of parts with my daughter and one of her friends and started building. Next thing I know, it’s almost midnight, and I’m plugging a microcontroller into my laptop. I asked Claude Code to figure everything out. And it did. It even made a little app that ran on wifi to control the robot from my phone. --- ## The Kit A week ago I ordered the **ACEBOTT QD001 Smart Car Starter Kit.** It’s an ESP32-based robot with Mecanum wheels (the ones that let it drive sideways). It comes with an ultrasonic distance sensor, a servo for panning the sensor head, line-following sensors, and an IR remote. It’s meant for kids aged 10+, but I’m a noob, soooo... whatever, I had a ton of fun! ## What Wasn’t in the Box Batteries. Apparently there are shipping restrictions for lithium ion batteries, so the kit doesn’t include them. If you want to do this yourself make sure to grab yourself the following: - **2x 18650 button-top rechargeable batteries** (3.7V, protected) - **1x CR2025 coin cell** (for the IR remote) - **1x 18650 charger** **A warning from experience:** NEBO brand 18650 batteries have a built-in USB-C charging port on the top cap that adds just enough length to prevent them from fitting in the kit’s battery holder. Get standard protected button-top cells like Nuon. Those worked well. You can get both at Batteries Plus. *One 18650 cell in, one to go. You can see here why the flat head screws were used to mount the power supply instead of the round head screws.* ## Assembly ACEBOTT had all the instructions we needed online. They have YouTube videos, but I just worked with the pdf. For a focused builder, this would probably take around an hour. For a builder with ADHD and a kiddo, it took around four hours. Be sure to pay close attention to the orientation of things. I accidentally assembled one of the Mecanum wheel motors with the stabilizing screws facing the wrong way. I had to take it apart and make sure they wouldn’t get in the way. *This is the right way. Flat heads don’t interfere with the chassis.* *Thought I lost a screw. Turns out the motors have magnets. Found it stuck to the gearbox.* *Tweezers were a lifesaver for routing wires through the channels.* *The start of wiring. Every module plugs in with a 3-pin connector — signal, voltage, ground.* *Couldn’t connect the Dupont wires at first — this connector pin had bent out of position. Had to bend it back carefully.* *Some of the assembly required creative tool angles.* *The ultrasonic sensor bracket. It looks like a cat. This was not planned. It’s now part of the personality.* ## Where Claude Code Jumped In Before I go too much further, I’ll just say that it would have been much easier if I’d given Ash the spec manual from the beginning. You’ll see why later. The kit comes with its own block-programming environment called ACECode, and a phone app for driving the car. You flash their firmware, connect to their app, and drive the car around. But we skipped all of that. Instead, I plugged the ESP32 directly into my laptop (after triple-checking the wiring) and told my locally harnessed Claude Code, we’ll call them Ash from here on out, to inspect the entire build and talk to it. *The ACEBOTT ESP32 Car Shield V1.1. Every pin labeled — but good luck figuring out how the motors work from this alone.* *All the wiring and labeling. What does it all mean? I've started plugging that back in to Claude and Gemini to learn more.* **Step 1: Hello World (5 minutes)** Within a few minutes, Ash wrote a simple sketch that blinked the onboard LED and printed the chip information over serial. It compiled the code, flashed it to the ESP32, and read the response. It did all of this from the CLI, the command-line interface. We didn’t use the Arduino IDE GUI at all. The ESP32 reported back: dual-core processor at 240MHz, 4MB flash, 334KB free memory. Ash got in and flashed one of the blue LED’s to show me it was in and reading the hardware appropriately. NOTE: I wish I’d waited to let my kiddo do more of this with me along the way. I got excited and stayed up to midnight working on it, but I should have waited. I’m going to make sure she’s more in the driver’s seat from here on out. *First sign of life. The blue LED blinking means Ash is in and talking to the hardware.* **Step 2: The Motor Mystery (45 minutes)** This next bit was my favorite because we had to work together to figure it out. Even though Ash was in, they had no good way of knowing which pins correlated with which wheel, nor which command spun the wheel forward or backwards. Ash figured out there were four motors but didn’t know which pins controlled them. The assembly manual listed sensor pins but not motor pins, and ACEBOTT’s website was mostly

Manual-Driven Development: 190 Findings, 7 Hours, Zero Rule Violations

https://preview.redd.it/lrsipzxvg5og1.jpg?width=2752&format=pjpg&auto=webp&s=eb9621bd6a89962f8abbe85cc87f6991dbea7717 Every Claude Code session you have ever had started with Claude not knowing your system. It read a few files, inferred patterns, and started coding based on assumptions. At small scale that works fine. At production scale it produces confident, wrong code, and you do not find out until something breaks in a way that tests cannot catch, because Claude wrote the tests against its own assumptions too. I call this confident divergence. It is the problem nobody in the AI tooling space is naming correctly. And it is the one that kills production codebases. Manual-Driven Development fixes it. Here is what that looks like in production numbers: Seven sections audited. 190 findings. 876 new tests written. 7 hours and 48 minutes of actual Claude Code session time against an estimated 234 to 361 hours of human developer time. That is a 30 to 46x compression ratio, reproduced independently across every section of a production codebase with 200+ routes, 80+ models, and a daemon enforcement pipeline that converts network policies into live nftables rules on the host. And across all seven sections, not a single CLAUDE.md rule violated. Not one. That last number is the one that should stop you. Everyone who has used Claude Code for more than a week has written CLAUDE.md rules and watched Claude ignore them three tasks later. The model does not do this deliberately. It runs out of context budget to honor them. MDD fixes the budget problem, and the rules hold. RuleCatch, which monitors rule enforcement in real time, reported 60% fewer rule violations during the SwarmK build compared to sessions running without MDD. Same model, same rules, same codebase. The only variable was MDD. I am not going to ask you to take that on faith. The prompts that produced these results are published. The methodology is documented. The section-by-section data is in this article. Everything is reproducible. If you are already using GSD or Mem0, you do not have to stop. MDD is a different layer solving a different problem. All three run together without conflict. I will explain exactly how near the end. The Problem Nobody Is Naming Correctly When Claude Code produces wrong code at scale, the community tends to blame one of two things: context rot, where quality degrades as the session fills up, or session amnesia, where Claude forgets everything when the session ends. GSD was built to solve context rot. Mem0 and Claude-Mem were built to solve session amnesia. Both are real problems. Both tools are real solutions. But there is a third problem that neither tool addresses, and it is the one that produces confident divergence. Claude does not know your system. Not in the way you do. It reads a few files, infers patterns, and starts coding based on assumptions. At production scale, with 200+ routes, 50+ models, and business rules distributed across a codebase that took months to build, the inferences diverge from reality. Claude produces code that compiles, passes its own tests, and is confidently wrong. Here is what makes confident divergence so hard to catch: everything looks correct. The code runs. The tests pass. Claude wrote the tests against its own assumptions about what the system does, not against what the system actually does. The divergence only surfaces in production, when a real user hits the edge case Claude never knew existed. Here is what makes it so hard to prevent: the problem is not just that Claude does not know your system. It is that you cannot reliably narrate your system to Claude either. You built the whole thing. You know how operator scoping works, how the tier hierarchy enforces access, how tunnels allocate /30 subnets in the 10.99.x.0 range. You know all of it in theory. But when you sit down to write a prompt at 11pm, you will not remember to mention that operators are scoped to specific groups and cannot modify policies outside their assigned groups. You will forget that ROLE_HIERARCHY is defined in three different files. You will not think to tell Claude that base-tier policies are system-only and cannot be created via the API. You are not going to enumerate 200 routes worth of business rules in a prompt. Nobody can. So Claude guesses. And confident divergence happens. That is the problem MDD solves. Not context rot within a session. Not forgetting between sessions. The deeper problem of Claude not having explicit knowledge of your system in the first place. The Token Obsession Is Solving the Wrong Problem Before explaining MDD, it is worth naming something about the current tooling landscape, because the framing most tools use will make MDD seem like another entry in the same race. It is not. Every tool launched in the last twelve months leads with the same promise: fewer tokens, lower cost, faster responses. Mem0 claims 90% token reduction. Zep claims 90% latency reduction. GSD kee