Last updated October 2026
How I work with AI
Coding agents write a lot of my code. These are the tools I actually use, what each one is for, and the habits that keep the output shippable. It changes every few months, so this page does too.
The tools
- Claude CodeMain coding agent · Max 5x plan
- Where most of the code gets written. The Max plan gives enough room to use the strongest models (Fable and Opus) on real, multi-file work instead of rationing them.
- Antigravity + GeminiQuick questions and edits · yearly plan
- Fast for questions and small modifications, and with a detailed prompt it goes a lot further. The real advantage is cost: cheaper than Claude Code, so it takes the high-volume small work.
- Claude appResearch and prototyping
- For research before a project starts and for prototyping, including quick design prototypes I can react to before anything is built.
- Matt Pocock's agent skillsSpecs, issues, grilling
- The ones I use most: writing a spec, splitting it into issues, and having the agent grill a plan before any code exists. Issues live as markdown files in the repo, so the agent and I work from the same source — this site included.
- MCP serversLaravel MCP · Next.js MCP
- Connected when a task needs the agent to work with the running app, not just the files. I also build MCP servers: Odigix's new admin exposes its operations through Laravel MCP.
- ImagesAPI skill or ChatGPT batches
- Either a skill that calls an image API from inside the agent, or the agent writes the list of images a page needs and I hand that list to ChatGPT, which generates them in one go and returns a zip.
- VSCodium on Linux MintEditor
- Faster and smoother than VS Code on my machine. The agents do the heavy lifting; the editor just has to stay out of the way.
The habits
The best model changes, and so does the way you use it
Every few months a different model is the strongest, and the habits that worked last quarter don't always carry over. Some models need everything spelled out; with others, pointing at the right files is enough. I re-check this with each release instead of assuming.
Plan before code
Spec, then issues, then a grilling session where the agent pokes holes in the plan. Most expensive mistakes are cheaper to catch in a markdown file than in a pull request.
Match the model to the cost of being wrong
Gemini for quick questions and small edits. Claude Code with the strongest models for multi-file work and anything touching money, permissions or data.
I review what ships
Agents write a lot of the code; I decide what gets built and review every change before it goes out. Deleting data and pushing to production stay behind explicit approval.