How I use AI in May 2026

Posted on: May 24, 2026

5 months later after my previous post like this, I wanted to write another post seeing what my outlook is on AI.

How I use

Model: GPT-5.5, $20 OpenAI plan, stil don't hit limits. I previously used the paid Claude sub, but I've jumped back to GPT models specifically cause of the next point.
Harness: Codex App. When I saw the workflow of project categorized sidebar, I knew that was the UX I wanted, and the application works reasonably well. For small sessions I also may just use the Codex CLI.
Still fairly hands on and reading the majority of the code and the output of sessions with these coding agents.
Still not really using AI in apps that have AI features (Browsers, web apps, etc) outside apps specifically for AI. Also still not using AI for creative work (art, video) as well as writing.
I Removed sandboxes due to environment issues, where a agent tries to do something but the container does not have the command installed and I don't really like keep on rebuilding such.That said, I am still not using --yolo / --dangerously-skip-permissions modes. I still don't know how to protect against bad commands.

Weeknotes Collector mainly for initial UI, but backend is mainly handcoded.
TCG Tracker coded primarially with Claude Code and Codex. Claude handled some niche things (Tanstack, Convex) quite well. Very minimal AGENTS.md.

Pricing is changing in some places (GitHub Copilot prices changing in June 2026, Anthropic switching to API pricing for programatic use, 3x higher API price for Gemini 3.5 Flash), but I am still using a subscription for my primary usage. Also, almost all of the providers now have a middle $100 a month tier the base $20 and max $200.
I use minimal amount of skills, have not used any MCP servers, and minimal AGENTS.md that I don't use too much of these unless I see the agents flailing over and over again.
- Of note, hand writing a few sentences in AGENTS.md seems to work well compared to having the model generate a long AI generated AGENTS.md file.