Codex Prompt Engineering

Knowledge snapshot: 2026-02-20

Purpose

Use this skill to design or review prompts for Codex/GPT coding agents with focus on:

• quality and reliability
• token efficiency
• correct tool usage
• safe autonomy

Use When

• Writing or revising system prompts for coding agents
• Debugging weak agent behavior (stalling, verbosity, bad tool calls)
• Calibrating reasoning_effort (none|low|medium|high|xhigh)
• Defining tool orchestration and planning behaviors
• Building evaluation loops for prompt iteration

Core Rules

1. Calibrate reasoning effort

• none or omit for trivial formatting/retrieval
• medium default for interactive coding
• high/xhigh for complex autonomous tasks

2. Run end-to-end autonomously

• gather context -> plan -> implement -> test -> refine
• ask only for ambiguity, destructive actions, or major architecture trade-offs

3. Use correct tool contracts

• file edits: apply_patch
• shell execution: exec_command
• task tracking: update_plan with plan: [{step, status}]
• batch independent calls with multi_tool_use.parallel

4. Keep communication compact

• tiny change: 2-5 sentences
• medium change: <=6 bullets
• large change: per-file summary + rationale

5. Evaluate systematically

• Analyze -> Measure -> Improve -> Repeat
• keep graders and representative datasets

6. Apply security basics

• validate/sanitize untrusted inputs
• defend against direct/indirect prompt injection
• enforce least privilege for tools/data

Runtime Mapping Note

Some environments expose shell tools under different names. In this repository runtime:

• use exec_command (not shell_command)
• use update_plan schema: plan with step/status

File Guide

• reference.md: compact canonical guidance and checklists
• patterns.md: reusable patterns beyond prompt engineering
• examples.md: concise before/after examples with correct contracts

Invocation

Use codex-prompt-engineering to review this prompt for quality and token efficiency.