Micro-blogs on Unmesh Mali

Rethinking my AI Coding Workflow

Wed, 07 Jan 2026 00:00:00 -0500

Why Git Worktrees may not be as useful as I initially thought when it comes to coding with AI agents.

My previous/current AI workflow :

Start local project
Enable git (git init)
Build a simple working version and test it in main branch (mostly with Anthropic’s Claude Code tool)
For advanced features, I would create n number of worktrees (n = number of coding agents)
Get each of the AI coding agents to work on a feature/functionality of the project independently (e.g. Claude Code, Opencode coding agent with GLM 4.7 model, Gemini GLI from Google)
Test each worktree locally independently and whichever is the best implementation - merge that one to main.

I am quitting this workflow now. Because managing multiple agents is a lot of work. Especially during the planning phase and when debugging. I use the AskUserQuestion tool on Claude Code - this tool triggers an interview where Claude Code asks me 15-25 questions about the new feature. Reading those, thinking about them, and answering/choosing the right option is satisfying but exhaustive. Other coding tools will also adopt this feature (check my last post about my ai_workflow) soon and I’ll end up duplicating the same process over and over. No thank you! Also Getting each of the coding tools (and their LLMs) to do the same thing burns through tokens like Confettu at New Year’s Eve on NYC Times Square. And not all LLMs or agents are good at every step of the process - each of them shine while doing different things (Claude Opus 4.5 shines with planning, Gemini 3 Pro on UI/UX, GLM 4.7 on general Python)- much like us mortal humans.

My AI workflow

Sun, 04 Jan 2026 00:00:00 -0500

Over the holidays, I have a lot of time to engage with technical novelties. And CLI coding was one of them. What triggered me trying to use CLI was Z ai’s new model launch - GLM 4.7. I snatched their $3 monthly coding plan and started coding via opencode. Why opencode? - because I already have a claude code subscription and use Claude models in it. And there was no easy way to use 2 claude code instances with Anthropic and GLM models. Hence opencode. As I worked through my Multi-modal AI app using these models, I got good results from GLM 4.7 and decided to make opencode with GLM coding plan a permanent tool in my AI coding arsenal.

Standing Dopamine

Mon, 27 Jan 2025 00:00:00 -0500

I have been struggling with limiting the screen-on time on my phone. I am addicted to instagram reels, YouTube, and especially Twitter. These website (with exception of YT) are highly negative and toxic these days for me - I have tried to tune the X homepage and it has helped slightly. But it is still very negative - maybe reflects the true thinking of the world, idk.

Anyway, I want to reduce the usage. I have tried many techniques and apps (like app timer) but they haven’t helped. I just go past those barriers and still use these apps. Today, a new idea hit me.