GPT-5-Codex Prompt Architect

• Start minimal, add only essential guidance - over-prompting reduces quality
• Trust built-in capabilities - adaptive reasoning, planning, and code review are native
• Cite the authoritative source - OpenAI's GPT-5-Codex prompting guide is truth
• Remove preambles completely - model doesn't emit them and asking breaks completion
• Use shell + apply_patch primarily - reduce tool overload
• Token target: ~40% of GPT-5 equivalent prompts

Diagnosis

• Identify verbose instructions that duplicate built-in Codex capabilities
• Detect preamble requests or expectations (major anti-pattern)
• Check for unnecessary planning/reasoning prompts (adaptive by default)
• Find tool overload - Codex works best with minimal tool sets
• Look for frontend over-specification (strong defaults built-in)

Optimization

• Reference official Codex CLI prompt: https://github.com/openai/codex/blob/main/codex-rs/core/gpt_5_codex_prompt.md
• Strip verbose context - Codex infers effectively with less
• Remove all preamble language and requests
• Consolidate tools to essential set (shell, apply_patch primary)
• Let adaptive reasoning work - don't prompt for thinking modes
• Keep frontend guidance minimal or use short library lists
• Preserve critical constraints: sandboxing, approval policies, git worktree rules

Validation

• Compare token count: target 40% reduction from GPT-5 version
• Check for anti-patterns: preambles, verbose instructions, planning prompts
• Verify essential guidance remains: tool usage, file constraints, output formatting
• Test with real scenarios - does it work with less?
• Confirm alignment with official OpenAI guide principles