Prompt of the Day: Make Your Terminal AI Agent Actually Useful

OpenAI Codex CLI is an open-source, terminal-native coding agent. Install it in 10 seconds:

npm i -g @openai/codex
# or
brew install --cask codex

It reads your working directory, plans tasks, and streams output directly in your terminal. No IDE required. Web search built in. Pipe support via codex exec --json.

But terminal AI agents have a failure mode: they go wide. You ask for one thing, they touch six files, refactor something you didn't ask them to refactor, and suddenly you're reviewing a diff that has nothing to do with your original request. Here are the three prompt patterns that fix this.

---

Pattern 1: The Scope Lock

Use this before any non-trivial task to lock the agent to exactly what you want:

• Only touch files I explicitly name, or files directly required by this task
• Do NOT refactor, rename, or reorganize anything outside the task scope
• If you identify something worth fixing that's out of scope, LIST it — don't fix it
• Maximum files changed: [N]

Success = task complete + no unrequested changes. ```

Why it works: The "list it, don't fix it" instruction is the key line. Agents love to fix things. This redirects that impulse into a useful artifact (a list of tech debt) without letting it derail your actual task.

---

Pattern 2: The Pipe-and-Confirm

For anything that writes, deletes, or modifies production files:

codex exec --json "[your task]" | tee /tmp/codex-plan.json

Then review /tmp/codex-plan.json before you run the actual execution. This gives you a dry-run view of what the agent plans to do — without it doing it. If the plan looks wrong, kill it before anything gets written.

For Codex CLI specifically: codex exec supports stdin piping, so you can also do:

cat error.log | codex exec "explain what's failing and suggest a fix — do not apply anything"

---

Pattern 3: The Daily Driver

For repeatable workflows, set a reusable system context in your shell config:

export CODEX_SYSTEM_PROMPT="You are a focused coding assistant. Before any task: (1) state what you will do in one sentence, (2) list files you will touch, (3) wait for me to say 'go'. Never begin executing until I confirm."

This turns every Codex session into a two-step flow: plan first, execute on confirmation. Adds 5 seconds per task. Saves you from 20-minute debugging sessions after a runaway agent.

---

So what? Terminal AI agents are fast and powerful, but they optimize for completion, not caution. These three patterns add deliberate checkpoints — scope lock before broad tasks, pipe-and-confirm before writes, plan-before-execute as a default habit. Takes 2 minutes to set up. Works with any terminal LLM agent, not just Codex CLI.

*Tested on Codex CLI v0.1.x and Claude Code. Both responded correctly to all three patterns on first run.*