Codex Skill Guide

Language

All prompts passed to codex MUST be in English.

Ask the user (via AskUserQuestion) which model and reasoning effort in a single prompt with two questions.

Model	Characteristics
`gpt-5.2-codex`	Latest frontier agentic coding model (current default)
`gpt-5.1-codex-max`	Codex-optimized flagship for deep and fast reasoning
`gpt-5.2`	Latest frontier model with improvements across knowledge, reasoning and coding

Select sandbox mode; default to --sandbox read-only unless edits or network access are necessary.
Assemble command with options (always include --skip-git-repo-check):
- -m, --model <MODEL> / --config model_reasoning_effort="<medium|high|xhigh>"
- --sandbox <read-only|workspace-write|danger-full-access> / --full-auto / -C <DIR>
Resume: cat /tmp/codex_prompt.txt | codex exec --skip-git-repo-check resume --last 2>/dev/null. If user requests different model/reasoning, insert flags between exec and resume.
Append 2>/dev/null to suppress thinking tokens (stderr). Show stderr only for debugging.
Run command, summarize outcome, inform user: "Resume anytime with 'codex resume'."

Use case	Command pattern
Read-only analysis	`codex exec --skip-git-repo-check -m MODEL --config ... --sandbox read-only 2>/dev/null`
Apply edits	`... --sandbox workspace-write --full-auto 2>/dev/null`
Network access	`... --sandbox danger-full-access --full-auto 2>/dev/null`
Resume session	`cat FILE \| codex exec --skip-git-repo-check resume --last 2>/dev/null`
Different dir	Add `-C <DIR>` to any command

After codex completes, use AskUserQuestion to confirm next steps. Restate model/reasoning/sandbox when proposing actions.

Stop and report failures whenever codex --version or a codex exec command exits non-zero; request direction before retrying.
Before you use high-impact flags (--full-auto, --sandbox danger-full-access, --skip-git-repo-check) ask the user for permission using AskUserQuestion unless it was already given.
When output includes warnings or partial results, summarize them and ask how to adjust using AskUserQuestion.

Three OpenAI prompting guides are available as Jupyter notebooks. Read the appropriate notebook when detailed guidance is needed.

User Question Type	Read This File
GPT-5.2 prompting, verbosity control, scope discipline, migration	`references/gpt-5-2_prompting_guide.ipynb`
GPT-5.1 agentic steerability, metaprompting, solution_persistence	`references/gpt-5-1_prompting_guide.ipynb`
Codex-Max starter prompt, tools (apply_patch, shell), compaction	`references/gpt-5-1-codex-max_prompting_guide.ipynb`

File: references/gpt-5-2_prompting_guide.ipynb

Key sections (grep patterns for navigation):

Key behavioral differences - Changes from GPT-5.1: deliberate scaffolding, lower verbosity, stronger instruction adherence
Controlling verbosity - Length constraints via <output_verbosity_spec> block
Preventing Scope drift - Prevent feature creep via <design_and_scope_constraints> block
Long-context - Handle 10k+ tokens via <long_context_handling> block
uncertainty_and_ambiguity - Hallucination prevention, clarification question patterns
Compaction - Extended context management via /responses/compact endpoint
Prompt Migration Guide - reasoning_effort mapping table (read reference for details)

File: references/gpt-5-1_prompting_guide.ipynb

Key sections (grep patterns for navigation):

File: references/gpt-5-1-codex-max_prompting_guide.ipynb

Key sections (grep patterns for navigation):

Clarify first: Use AskUserQuestion for ambiguous requests before crafting prompts
Structure the prompt: Frame tasks as "complete end-to-end", include context, define constraints explicitly
Execute and iterate: Use metaprompting (root-cause analysis + surgical refinement) rather than complete rewrites