A Quiet tmux Desk for Cheaper AI Coding Sessions

A Quiet tmux Desk for Cheaper AI Coding Sessions

The expensive part of a long coding session is not always the model. It is the loop: lose your place, reopen three tabs, ask an assistant to reconstruct the problem, then pay for another oversized conversation. My terminal setup is designed to make that loop harder to enter. Fewer surfaces mean fewer interruptions, less repeated context, and fewer purchases disguised as productivity improvements.

One tmux Session, Two Windows, No Dashboard

Start with the terminal and editor you already have. Create a session with tmux new -s focus. Keep the editor in the first window and your shell in the second. Split the shell window only when you need a test runner beside an AI CLI. Three permanent log streams are usually three invitations to stop thinking.

I disable the tmux status bar with tmux set -g status off and keep window switching on keyboard shortcuts. If you rely on the bar to identify remote hosts, keep that information visible instead. A minimalist setup should never make production mistakes easier.

• Enable your operating system’s focus mode before opening the repository.
• Disable audible terminal bells and flashing activity notifications during focused work.
• Keep chat, mail, and browser dashboards outside this workspace entirely.
• Use one readable font instead of shopping for another terminal application.

A Five-Line Task Note Prevents Paid Reorientation

Before starting the assistant, write a small scratch note: the desired behavior, the failing test, the relevant files, the next action, and the stopping condition. Keep it in an ignored local file or outside the repository. Check git status before committing so private notes do not accidentally ship.

When I return from a break, that note restores the task faster than scrolling through an agent transcript. It also becomes a compact handoff when a conversation grows unwieldy. Start a fresh conversation at a genuine task boundary, not after every question: repeatedly sending the same background can create its own waste.

Small Diffs Beat Repository-Sized Prompts

Run the narrowest relevant test locally before asking for help. Capture the failing assertion and enough surrounding output to explain it. Do not paste an entire build log when the useful error occupies twelve lines. Likewise, ask the assistant to inspect specific files before authorizing broad exploration.

• Request a short plan before allowing edits across multiple directories.
• Review the diff before requesting another implementation or cleanup pass.
• Set a retry limit when an agent repeats unsuccessful commands.
• Exclude secrets and irrelevant generated files from the material you share.

One Relay Configuration, Deliberate Model Choices

Switching providers should not require another browser session. Fluxo gives you one key and one base URL, https://api.fluxorelay.com, for Claude, GPT, Gemini, DeepSeek, Kimi, GLM, Qwen, Grok, and more. It works with Claude Code, Codex, Gemini CLI, and any OpenAI SDK. Follow each client’s configuration instructions rather than assuming every CLI uses identical environment variables.

Keep credentials out of shell history and tracked configuration. Use your client’s supported credential storage, or inject secrets through your existing secret manager. Save model choices in configuration so switching tasks does not become another setup project.

For example, try Claude Haiku 4.5 on a bounded explanation or small test change, then evaluate Claude Sonnet 5.5 when a task needs deeper investigation. Measure acceptable results and total usage rather than assuming either choice always costs less. A cheaper attempt that triggers repeated repairs may be the expensive route.

A Session Budget Includes Cache Traffic

Fluxo has no monthly fee and bills per real token, including cache reads and writes. Do not treat cached context as free. Check usage after a representative task, keep stable instructions consistent where the client supports caching, and avoid padding prompts simply because caching exists.

Set a spending ceiling for the session and check usage at natural breaks. You can create a Fluxo account when you are ready to test this workflow. Users also earn a referral share when people they invite pay; treat that as incidental, not a reason to generate more traffic.

End by recording the next action and closing noisy processes. Tomorrow’s cheapest prompt is often the one your task note makes unnecessary.