codex - SKILL.md Agent Skill

name: codex description: Run code analysis, reviews, and consultations using Codex CLI (OpenAI). Executes safely in read-only sandbox mode. allowed-tools: - Bash - Read - AskUserQuestion user-invocable: true

Codex CLI Skill

Use Codex CLI (OpenAI) to perform code analysis, reviews, and consultations.

Triggers

Load this skill for the following keywords:

"codex", "ask codex", "consult codex"
"review this", "look at the code" (in code review context)
"architecture analysis", "design consultation"

Do NOT load for

Simple file reading (Read tool is sufficient)
Questions Claude can answer alone
Code editing execution (codex is used in read-only mode)

Command Syntax

codex exec --full-auto --sandbox <mode> --cd <dir> "<prompt>"

Parameters

Parameter	Value	Description
`--sandbox`	`read-only`	Read-only mode (for analysis/review) Default
	`network-off`	Network disabled (allows local changes)
	`off`	No restrictions (Do not use)
`--cd`	path	Target directory
Bash timeout	300000ms	Codex execution may take time

Prompt Structure

All prompts should follow this structure:

[Role definition (optional)]
[Task description]
[Scope/constraints (optional)]
[Output format (optional)]

No confirmation or questions needed. Output specific suggestions, fixes, and code examples proactively.

The closing instruction is mandatory. Omitting it causes Codex to enter interactive mode, resulting in incomplete output.

Workflow

Step 1: Understand the Request

Classify user request (review/bug investigation/design consultation/etc)
Use AskUserQuestion if unclear

Step 2: Confirm Target Directory

If specified: use that path
If unclear: use AskUserQuestion to confirm
Default: current working directory

Step 3: Determine Sandbox Mode

Default: read-only (for analysis/review)
Consider network-off only if changes are needed
Never use off (security risk)

Step 4: Compose Prompt

Select appropriate template
Incorporate user's specific concerns
Always include closing instruction

Step 5: Execute Command

Run codex exec using Bash tool
Set timeout: 300000ms (5 minutes)

Step 6: Report Results

Summarize and structure lengthy output
Prioritize important findings by severity
Suggest follow-up actions if needed

Use Case Templates

1. Code Review

codex exec --full-auto --sandbox read-only --cd <dir> \
"Review the code in this project.
Aspects: code quality, readability, maintainability, potential bugs, security
Output format: Classify issues by severity (High/Medium/Low) with fix suggestions for each

No confirmation or questions needed. Output specific suggestions, fixes, and code examples proactively."

2. Bug Investigation

codex exec --full-auto --sandbox read-only --cd <dir> \
"[Symptom description]

Investigate the cause and output:
1. Cause identification (root cause and direct cause)
2. Impact scope
3. Fix proposal (with code examples)
4. Prevention measures

No confirmation or questions needed. Output specific suggestions, fixes, and code examples proactively."

3. Refactoring

codex exec --full-auto --sandbox read-only --cd <dir> \
"Identify technical debt in this codebase and create a refactoring plan.
Prioritize by impact × cost and include Before/After code examples for each item.

No confirmation or questions needed. Output specific suggestions, fixes, and code examples proactively."

4. Architecture Analysis

codex exec --full-auto --sandbox read-only --cd <dir> \
"Analyze the architecture of this project.
Output: (1) Current structure diagram (Mermaid), (2) Dependency issues, (3) Improvement proposals

No confirmation or questions needed. Output specific suggestions, fixes, and code examples proactively."

5. UI/UX Evaluation

codex exec --full-auto --sandbox read-only --cd <dir> \
"Evaluate this project as a UI designer.
Aspects: visual hierarchy, spacing, color contrast, accessibility, cognitive load
Provide Tailwind CSS/CSS fix examples for each issue.

No confirmation or questions needed. Output specific suggestions, fixes, and code examples proactively."

6. Specific File/Function Analysis

codex exec --full-auto --sandbox read-only --cd <dir> \
"Analyze <function-name/class-name> in <file-path>.
[Analysis aspects or concerns]

No confirmation or questions needed. Output specific suggestions, fixes, and code examples proactively."

7. Copy/Message Review

codex exec --full-auto --sandbox read-only --cd <dir> \
"Review user-facing messages in this project (error messages, UI copy, notifications, etc.).
Aspects: clarity, consistency, tone, internationalization
Provide specific improvement suggestions.

No confirmation or questions needed. Output specific suggestions, fixes, and code examples proactively."

Error Handling

Error	Resolution
`codex: command not found`	Guide installation: `npm install -g @openai/codex`
API authentication error	Advise checking `OPENAI_API_KEY` configuration
Timeout	Narrow scope (specify directory/file)
Empty/incomplete output	Verify closing instruction in prompt, suggest retry

Security Considerations

Sandbox required: Always use --sandbox read-only
Sensitive data: Warn when directory contains .env, credentials, or API keys
Output review: Verify codex output contains no sensitive information before reporting

Notes

For large repositories, narrow the scope (specify directory or file)
Maintain --sandbox read-only for repositories containing sensitive information
Summarize and structure lengthy output for the user