Nick

I run a team of Grok Bots that ship real apps. I write almost none of the code now. 11 PRs in about 5 days with 3 bots. Tests went from 9 to 70. 140M tokens. Here's exactly how it's set up 🧵

Working with small businesses, I realized very quickly they have too many dashboards and no place to just ask a bot/agent for help. So my first task is always create one for them to answer a simple question such as: "who hasn't paid?" and it gets their answer right away.

I had been in a rut and lost my love for developing software outside of work, but in August I watched a video of @poteto discuss the new @cursor_ai and Grok Bot. It got me excited to develop again and haven’t stopped since.

Every weekly reset is like payday. I start running my heavy builds and updates right away, and hopefully have some leftover for the fun stuff.

One thing I’m loving about the SpaceXSI unified account is having a bot try my app before users do. It walks through as a brand-new user and reports what sucks. That's the cheapest alpha tester there is.

My view: small business don't want another pretty dashboard. They just want a simple agent or routine to handle listings, replies, and order busywork, and keep a person approving or answering to things that matter. That split is what makes it valuable.

I’ve found that the second time I explain a task to a Grok Bot, I have it save that as a skill because a long chat history can get lost in the chaos. Skills are what keep it consistent over time.

Happy Halloween. The scariest words in AI development: "the agent says it's done." To save myself from a headache, I handle all PRs. That way mistakes are still my fault. Agent done = My turn to work.

I’ve been working to write each Grok Bot's description like a job posting for a coworker or new employee. Clear concise instructions with a narrow job so it can work without getting distracted.

Review the walkthrough before the diff. Lately I’ve caught an agent's screenshots or video show the wrong thing, and there's no reason to read all of that new code for the PR if it didn’t get done right.

I never start with a blank canvas for new Grok Bots. I take my best bot's standing rules, paste them in, and delete whatever I don’t need. Fixing rules is faster than inventing them.

Even with thousands of Grok Bots in the directory, small businesses don’t need to access that many. It needs one or two bots doing boring jobs every day, like turning order emails into a daily list or helping to not miss phone calls. One job done daily beats a thousand bots you never need or open.

An investor can run 20 Grok Bots at his firm to research and make trades. I run a handful to develop and ship apps. It’s the same playbook either way, bot count doesn't matter. What matters is whether each one has a one-line job it can't wander away from.

For a small business, the hard part isn't the coding agent. It's picking which manual process to kill first. I start with either the task somebody hates doing, or the biggest road block holding up the most work. Start there and the tool sells itself.

Every Grok Bot I create gets two lists: a job description and an anti-jobs list. The job tells it what to do. The anti-jobs list (never post, never touch the code, never message anyone) is what keeps it trustworthy. The second list matters more in my opinion.