AI Agents: How I Use Them

Harness-agnostic setup and principles live here; tool-specific tips live on each tool’s own pages. This is a living document that evolves with my workflow.

Harnesses#

• Claude Code: how I use it, tips, skills
• OpenAI Codex: terminal TUI and app.
• Other harnesses and multiplexers: see the alternatives list

Environment#

• Agent interface (comparison of the options):
  • Paseo (self-hosted, multi-agent app)
  • OpenAI Codex app
  • Ghostty + Herdr
• Dictation: Superwhisper (running a local, on-device Parakeet model) for voice input. My global instruction file has a note about interpreting phonetic typos.
• Config managed via dotfiles repo (~/.claude/ symlinked into ~/git/dotfiles/claude/)

I currently run agent sessions in Paseo that lets me use a Codex-app-style interface on Claude Code and other agents. A small but very useful feature: explicit control over when a follow-up prompt is injected—send it now, or queue it for the agent’s next turn. This is what Codex offers but not by Claude Code. In Claude Code tui, you can’t predict when a typed prompt will be injected.

I may try both Herdr (an agent multiplexer) inside Ghostty and Paseo for a while. Earlier I tried Orca, which is powerful but a bit too complex for my taste (its whole model is built around Git worktrees), and cmux, which is lightweight and similar in spirit to the Ghostty + Herdr combo but had memory-leak issues. Earlier still: bare Ghostty splits (⌘+D / ⌘+⇧+D), and before that tmux (see running multiple sessions).

When I work in the terminal, my default layout is two panes side by side—the agent on the left, Neovim on the right. I split further as needed but try to return to this two-pane setup. Neovim is great here for the same reason as Ghostty: snappy and versatile. You can quickly move around, explore files, and make edits while the agent works in the other pane, without the need of a separate, heavy editor app. I sometimes run watch scripts (e.g., auto-compile, file sync) in a tiny extra pane.

Usually I have several tabs open (one for each project) with each tab having 2+ panes.

Principles for Context engineering#

Organize instructions with progressive disclosure—see Context engineering. Don’t over-engineer the setup, and invest in verification; both are discussed on the Claude Code page but apply to any harness.