AI tooling changes constantly, so this page is a snapshot of how I work right now. I’ll keep it updated as the workflow changes.
My current workflow
I built my own task manager, the orchestrator behind kstack/k-os. It’s local-first and tmux-native: an AI task runner.
It works in stages:
- Planning. A planning agent turns my ideas into tasks. Each task has acceptance criteria and dependencies.
- Research first. Investigation tasks produce a summary doc. I review it before any implementation tasks are created.
- Execution. An orchestrator runs every task whose dependencies are done. Each one gets its own tmux window, and the loop keeps going until everything is done.
- Review. Some tasks end in a “review” step, so I check the results before anything is marked done.
Claude Code is the underlying agent. I give each task an opinionated agent profile (python-dev, unit-test, code-reviewer, and so on) and draw on reusable skills for the rest. The code itself is private, but I’d encourage everyone to build their own.
Tools I’ve tried
- OpenCode AI: terminal harness similar to Claude Code, and it can use local LLMs.
- Planning With Files: Markdown plans lead to great results. I’ve been doing this for most of the year.
- Superpowers: opinionated agent skills and configuration that improve performance.
- Oh My Opencode: similar opinionated agent configuration, still exploring.
- SuperWhisper: local speech to text.