In the last 30 days, I burned through 2.2M tokens on Claude + 2M on Codex in my spare time and built 5 projects: an iOS app in Swift, an Android app in React Native, multiple web apps with backends, etc.
This has been the most fun I've had building in years. The only real limit right now is imagination and not getting distracted.
Here are my top takeaways for building with AI agents in early 2026:
1/ I basically never open my IDE anymore (except to review diffs)
100% of coding happens via agents. I talk to the model using voice (Wispr Flow), start almost exclusively in Plan mode, and let it iterate. Once you learn each model's quirks and prompting style, the speed is insane.
For example, Codex 5.2 Max is amazing for single-shot but it reads your code for like 10 minutes before executing. I prefer Opus for day-to-day mostly because my tooling is better setup for Claude Code.
2/ Test-driven development is back and the only way to keep your sanity
Near-100% test coverage gives me real confidence in AI-generated code. I bake instructions into my project docs/agents.md: "Whenever new functionality is added, cover critical business logic with tests." When debugging, I tell it to start with failing tests. This is the primary way to build confidence in the code the agent ships.
3/ Skills are the new DRY ("don't-repeat-yourself") principle
Anything I find myself prompting repeatedly, I turn it into a reusable skill (and usually have the model write the skill itself). I literally ask it "Build a skill so you can do what you just did better and faster next time". This compounds fast.
4/ Use the agent to improve the agent
When it goes off the rails, I literally ask: "What do you wish you knew to give a better response?" Recursive self-improvement is one of the biggest unlocks.
5/ Keep CLAUDE.md / AGENTS.md updated automatically
Add a rule: "Whenever you make important changes, keep this doc updated." Pair it with a "keep it succinct" skill. My Claude.md docs stay surprisingly clean and useful.
Bonus workflow hacks
alias yolo → claude --allow-dangerously-skip-permissions. This lets me just go into yolo-mode when I want to bang something out and not require permissions.- I added a pre-commit cleanup prompt inspired by Steve Yegge's "landing the plane" idea. Before the model tells me it's "done", I ask it to do 2 things: (a) Use the code-simplifier Claude plugin to simplify the code it just created. (b) Use a code-review skill to review its code as a PR reviewer would. Then make the changes requested if they are deemed to be valid. This burns through more tokens but I get way better code outputted which I can actually review.
- For UI/browser testing, I'm using agent-browser and its excellent within Openclaw. However, locally I actually just use the Claude Chrome Extension and connect it with Claude Code.
- For parallelizing local work, multiple repo checkouts with pnpm feels more intuitive than git worktrees for me right now.
If you're building with frontier models in 2026, what are your biggest workflow wins or frustrations? Keep shipping.