AI Workbench
My AI Stack Is A Workbench
Why I do not treat models as favorites, but as tools with different jobs inside a larger workflow.
I don’t have a favorite model. I have a workbench.
Each model has a job.
Claude Code is where I start when the work is still fuzzy: discovery, design, specs, audits. Codex is where I hand a spec when I want something built fast. Gemini handles deep research when I need a wide pass across a topic. NotebookLM is for working inside source material once I have it. Local models like Qwen or Hermes run on a Mac mini for lightweight tasks when I want to save tokens.
The point isn’t to collect tools. The point is knowing what each one is actually good for.
That only comes from use. Gemini can loop. Claude Code is strong but expensive; session limits matter. Codex builds fast, but I don’t trust it for design work the way I trust Claude Code. Those aren’t complaints. They’re operating facts.
My rule: no model owns the final product.
The models draft, research, build, audit, and refactor. I make the decisions: structure, direction, quality. I steer. I decide when it’s right.
That matters because models are persuasive. Without thread discipline, you can spend hours chasing the next recommendation. Manage tokens. Watch for drift. Return to the original intent.
The workbench is valuable not because of the list of tools. It’s valuable because you know which tool belongs where, when to switch, and when to stop listening.