JBS Weekly
Most AI projects at small companies have the same problem: one person built it, that person is still the only one who can run it, and if they're out sick or leave, the whole thing stops.
This week I want to fix that. I'm walking through the five things that belong in every AI project folder, in order, so that someone who wasn't in the room can open it and actually figure out what's going on. No prerequisites. Just a structure that works.
🛠️ This Week’s Build
The Five-File Folder That Makes AI Projects Transferable
Most AI project folders are graveyards. Scripts with no context, prompts with no explanation, workflows nobody remembers building. Six months later, the person who built it is the only one who can touch it, and even they're guessing.
Here's the structure I use to fix that. Five components, in order.
First, a top-level memory file. Think of it as a README written for a very literal new employee. It covers what the project does, which tools it touches, and the two or three rules that keep getting broken when nobody writes them down. This is the file someone reads first. Everything else is secondary.
Second, situational instruction files. Not everything belongs in that top file, or it becomes a wall of text nobody reads. Pull the specific stuff, like how to handle a failed API call or how to deploy the workflow, into its own short file. One line in the main file points to it. Keep it navigable.
Third, a written definition of done, created before you build anything. One page: the goal, the constraints, what's explicitly out of scope. I've watched teams spend two weeks building an automation, then rebuild it entirely once someone finally asked what it was supposed to do. That's not a technical problem. That's a documentation problem.
Fourth, modular workflows. One workflow pulls data. One formats it. One sends it. Three simple pieces beat one workflow that does everything and breaks in three places every time something changes. The instinct is to keep it consolidated. The smarter move is to keep it separable.
Fifth, a testing environment that isn't the live system. A duplicate workflow, a staging copy, anything that isn't production. This is the one people skip until something breaks in front of a client.
The test for whether you've done this right is simple: someone who wasn't in the room can open the folder, read one file, and tell you what the project does and how to change it safely. If they can't do that, the folder isn't done yet.
📰 AI News This Week
Anthropic Publishes Agentic Misalignment Research
Anthropic released a report showing early signs of AI agents behaving unpredictably in simulated environments: sabotaging code, encouraging whistleblowing, acting outside their intended scope. This isn't a theoretical concern. It's documented behavior from agents that looked fine under normal conditions.
Joe's Read: If you're running or planning to run autonomous AI agents inside your business workflows, this is a reason to build in human checkpoints now, before you scale anything.
Thinking Machines Launches Inkling, a Smaller Targeted AI Model
Mira Murati's new company, Thinking Machines Labs, released its first model called Inkling. It's built around smaller, task-specific architectures rather than trying to do everything. It handles text, images, and audio.
Joe's Read: Ops teams running narrow, repeatable AI tasks, like document processing or intake triage, may get better results from a focused model like this than from a generalist one.
Walden Robotics Raises $1.1B to Put Humanoid Robots on Factory Floors
Walden Robotics, a Toyota spinoff less than a year old, raised $300 million at a $1.1 billion valuation. Their robots already work full shifts on a Toyota production line. They use wheeled bases with humanoid upper bodies and learn new tasks through hands-on practice using what they call Large Behavior Models.
Joe's Read: This is most relevant to operations leaders in manufacturing or warehouse environments who are tracking when physical automation becomes a realistic near-term capital decision.
🧰 Tool Worth Trying This Week
Agently
Agently is built around creating a digital brain for your company and using AI agents to automate workflows across it. The pitch is that you connect your business knowledge and processes, and agents handle execution without you managing each step manually.
This is not a beginner tool. If your processes aren't documented and your data isn't organized, plugging agents into the chaos won't help. Get your foundation in order before you go anywhere near this.
🗺️ From The Field
The thing that breaks AI projects isn't the automation itself. It's the assumption that whoever built it will always be available to explain it.
Every project I've seen fail the handoff test failed for the same reason: the builder kept the important context in their head instead of in a file. The workflow ran fine as long as they were around. The moment someone else had to touch it, everything slowed down.
Documentation isn't a cleanup task you do at the end. It's a design decision you make at the start. The folder structure I described this week forces that decision early, before the build gets complicated enough that writing it down feels like too much work.
If your current AI projects only work because a specific person is babysitting them, that's not automation. That's a dependency you dressed up as a system.
🤔 Joe’s Take
Back in 2023, right as ChatGPT was taking off, I built an automation that had nothing to do with AI. Just RPA and some scripting, but it solved a real problem. My team was manually processing 15 to 25 entries a day, each one taking about 20 minutes, and human error kept creeping in: missed appointment times, wrong ID numbers, bad addresses.
I built a form with logic that locked in the mission-critical fields so they were always standardized and correct. It worked. That automation was saving the company about $20,000 a month in operational cost.
Then my role got eliminated, out of nowhere, and I never documented how I'd built it. The moment they shut off my credentials, the whole thing stopped working. Someone could have rebuilt it eventually, but in the meantime the team went from a system that ran itself back to doing it all by hand, like trading a dishwasher for a bucket and a hand pump.
The lesson isn't subtle. If you're the only one who understands how something works, you haven't built a system. You've built a dependency. And the day you leave is the day the company finds out the hard way.
⚒️ Tools I Use
n8n — The automation tool I use to connect apps, trigger workflows, and stop doing things manually. If there's a repetitive process in your business, this is where you start fixing it.
VoiceInk — A local AI dictation tool for Mac that transcribes your voice with near-perfect accuracy and runs entirely on your device, meaning nothing you say ever touches a cloud server.
Blotato — Handles the full content distribution side of your business: drop in a topic and it generates platform-specific posts, or feed it existing content and it repurposes it across formats. TikTok videos become tweets, podcasts become blog posts. Includes a scheduling calendar, visual creation tools for carousels and infographics, and publishes natively to 9 platforms with no per-post fees.
Beehiiv — What you're reading right now is published on Beehiiv. If you're thinking about starting a newsletter or moving off a clunky platform, this is the one I'd recommend. 20% off your first 3 months with my link.
Google Workspace — Beyond email and Docs, a Business Standard plan includes Gemini Pro built into every app, NotebookLM Plus, and access to the enterprise versions of the whole suite. Better value than a standalone Gemini subscription when you're already paying for Google anyway. 14-day trial and 10% off your first year.
Descript — Video and podcast editing that works like a text document. You edit the transcript and the media follows. Cuts filler words, cleans up audio, and handles captions automatically. 50% off your first two months on the Creator Plan.
💭 Final Thoughts
A workflow that only one person can run isn't a system. It's a single point of failure with a nice UI. The folder structure I walked through this week isn't about being tidy. It's about making your AI projects transferable, maintainable, and survivable when the person who built them isn't in the room. Start with the README. Everything else follows from that.
PS: If you want your existing AI projects documented and structured so someone else on your team can actually own them, book a discovery call and we'll build the solution together in a live session.
Cheers,
Joe