So I've been asked a few times what people think of vibe coding their own software, so I really wanted to share this one here. I think it's a fair thing to try. You can get an AI to build you a working app in a weekend, and by Sunday night it looks finished. The part nobody mentions is what comes after, so I sort it into three questions: 1. What does it touch? Client details, your prices, your files, and what the AI is allowed to change. A 2026 study on arXiv audited 200 deployed apps built with AI coding tools and 91% had at least one vulnerability. They were open source apps built with two tools and the authors say it may not generalise, so I read it as a reason to test, not a verdict. 2. How does it behave? In your own software the AI only gets what you give it. Write down what it may do, what it must never touch and where it stops and asks a person. Then test it on ten past cases and rerun them every time you change the prompt or the model. 3. What does it cost to keep? Providers retire models, and hosting and testing services bill by usage with defaults that can catch you out. My rule of thumb: build it yourself if it's internal, holds no client details and moves no money. Check it first if it touches client records or payments. Hand it off if you can't do those checks or nobody can say who looks after it next year. I put the full walkthrough in a video if you want it: https://youtu.be/uBqzgb07_8U

Grok Bot now handles most of my day-to-day admin. I still use Claude for structured work like a detailed estimate, where I want tested skills and control over the output. For the small random jobs, Grok Bot is simpler because it writes its own skills, keeps its own memory and sets up its own scheduled tasks. All you manage is the bots. Two rules get the most out of it. First, if you're not sure it can do something, ask it. When it told me it couldn't connect to SharePoint, I asked for a workaround and it found one. Second, give every bot one clear goal. A bot doing ten different jobs gets its memory muddled. Start by telling it about your business and connecting your software. Use an MCP or API connection where one exists, and a browser login as the slower fallback. Give anything sensitive read-only or draft-only access. - Business development: finds tenders and gives a go/no-go recommendation - Pre-construction: tracks bids and follows up submissions - Commercial manager: protects margin and contract position - Finance: cash flow and getting paid - Systems and compliance: insurances, SWMS and paperwork - Project assistant: one per project, coordinating delivery A general manager bot can sit on top and hand work to the others. Which job would you hand to a bot first? Full walkthrough: https://youtu.be/0hzqCMowrys

Most construction businesses run on Microsoft, so Copilot is often the AI you already have. It's more confusing than ChatGPT or Claude, but it makes sense once you treat it as three separate tools. Chat is for questions and answers. Turn on Work IQ so it can read your SharePoint files, email and Teams, and set up one notebook per project so each chat has the right context. Agents are reusable instruction packs: you open the agent, then tag the project notebook into the chat. That's the reverse of how projects work in Claude or ChatGPT. Co-work is where Copilot edits files and runs tasks inside a SharePoint project. You can schedule jobs like a daily project update in Teams, load skills and connect other apps. Custom connections usually need your IT admin's approval. Copilot Studio is for building company tools, like more advanced agents, fixed workflows similar to Zapier, and simple apps. Agents you build there show up in Chat, but I still can't get them to appear in Co-work. If your company is on Microsoft 365, which of these three do you actually use? Full walkthrough: https://youtu.be/rQynvHYzqNo

I've moved from Claude Cowork back to Claude Code for nearly all my work. You open it on a folder on your desktop, type your prompts and it does the work. Cowork now runs in the cloud on copies of your files, so you can't move or delete the originals which is super frustrating. These are the things I've been using: - Side chats (/btw) for quick questions without cluttering the main chat. - /goal for long jobs, like checking every contract clause against your standard terms or categorizing 100 invoices. - Sidebar sections, so each project keeps its chats together. - The session rail, which shows how much context you've used. I set auto-compact at 400,000 tokens because results get worse as the window fills. I also built a take-off checker with Mods that flags anything Claude was unsure about, but I'm still working out where Mods are most useful. The idea seems cool, just unsure how useful it actually is Full walkthrough: https://youtu.be/fs4PGGZmp_0

For me, it’s poorly prepared bid packages. You get sent hundreds of documents that all say different things. Then you spend 10 hours going through them to work out what the client wants. Unfortunately AI is making it even easier to produce hundreds of pages of junk for someone else to read. What’s yours? What part of working in construction do you wish someone would sort out? I’ve been working on a new product. And yes, this is one of those annoying product research surveys where I ask what your problems are and how I can help. If you’ve got five minutes, I’d really appreciate you filling it out. https://docs.google.com/forms/d/e/1FAIpQLSdILOUFSeoYJRJOfY5g5BHnI8ZLPHcfZhiL78nOE0AFOpRdTQ/viewform?usp=header