| name | claude-codex-skill-toolkit |
| description | Convert, install, audit, package, and troubleshoot Claude Code skills for Codex. Use when the user asks to install a Claude skill in Codex, migrate a Claude-oriented skill, adapt a GitHub skill repo, fix Codex skill discovery, package a migrated skill for sharing, or review whether a Claude skill will work well in Codex. |
Claude to Codex Skill Toolkit
Purpose
Handle the full Claude-to-Codex skill lifecycle:
- inspect a GitHub repo or local skill folder
- inspect shell snippets that manually create
.claude/skills/<name>/skill.md
- choose the shortest safe route: direct install or migration
- classify the install hint the user found in a Claude tutorial
- score compatibility, dependency risk, and install tier before acting
- rewrite discovery-critical metadata when needed
- validate that Codex can actually discover the result
- package the finished skill for reuse or GitHub publishing
Use this skill for both one-off installs and building a reusable, polished Codex-native port.
Auto Trigger
Activate when the request includes ideas like:
- install this Claude skill in Codex
- migrate this Claude Code skill
- port this skill to Codex
- adapt this skill for Codex
- will this mkdir/printf
.claude/skills/.../skill.md command really install a skill
- why is this installed skill not showing up
- package this migrated skill for GitHub
- review whether this Claude skill will work in Codex
Treat the following as direct Claude-oriented signals even if the user only says "install this skill" and pastes a link:
- GitHub URLs or
owner/repo references under anthropics/, especially anthropics/skills
- URLs, repo names, or text that mention
Claude, Claude Code, Anthropic, or anthropics
- paths or snippets containing
.claude/skills, .claude-plugin, skill.json, or marketplace.json
- commands such as
/plugin marketplace add, /plugin install, or /plugin add
- Claude-specific ZIP upload flows or Skills API references
When any of those markers are present, use this skill first before relying on raw direct install.
Quick Route Selection
Choose the shortest route that is still reliable.
- If the repo already contains a clean
SKILL.md and the folder layout is obvious, install the exact skill folder first, then validate discovery.
- If the repo contains platform-specific layouts such as
.claude/skills, .claude-plugin, or skill.json, inspect those signals before deciding it is non-installable.
- If the user provides a shell snippet that creates
.claude/skills/<name>/skill.md, treat it as a local scaffold and inspect the generated folder before calling it installed.
- After any direct install, run the Codex-native polish checklist. A valid
SKILL.md can still contain Claude-only behavior.
- If the repo is Claude-specific, ambiguous, multi-skill, or still uses Claude-only assumptions, migrate before claiming success.
- If the task is only evaluation, inspect the repo and explain compatibility without installing anything.
- If the user wants a shareable artifact, build a clean Codex-native package with minimal structure and validation notes.
Use scripts/inspect_skill_repo.py when a local folder is available and the repo layout is not obvious.
Use scripts/check_skill_md.py when you need a deterministic check of frontmatter or BOM issues.
Use scripts/classify_install_hint.py when the user pastes a Claude tutorial command or install snippet and you need to map it to a Codex route.
Use scripts/trigger_score.py when installability looks fine but you need to know whether the skill will actually match real user prompts well.
Use scripts/generate_install_report.py to leave behind a readable Markdown analysis after each install review.
Use scripts/install_skill_flow.py when you want a semi-automatic route selector that can also copy a chosen local skill into Codex.
Core Workflow
-
Inspect the source.
Read the repo layout, locate skills/, SKILL.md, helper scripts, and any Claude-only packaging assumptions.
-
Decide the route.
Use references/compatibility-matrix.md to decide between direct install, light adaptation, or full migration.
If the user copied a tutorial command, normalize it first with references/install-route-adaptation.md.
-
Perform the smallest viable change.
- Direct install if the source is already Codex-friendly.
- Rewrite
SKILL.md and supporting files if discovery or trigger quality is weak.
- Remove or rewrite Claude-only instructions, plugin hooks, marketplace assumptions, and broken command examples.
- Split long guidance into
references/ when building a durable Codex port.
- Use compatibility score and install tier to explain why a route is safe or risky.
- Generate an install report whenever the result will be reviewed later, shared, or used as a migration record.
-
Polish for Codex-native behavior.
Use the checklist below and the longer playbook in references/migration-playbook.md. Preserve behavior, but rewrite packaging and operational guidance so the installed skill is usable by Codex.
-
Validate, do not assume.
Always verify:
- the skill folder exists where intended
SKILL.md is parser-friendly
- the file is UTF-8 without BOM
codex debug prompt-input exposes the skill when the target is a live Codex install
- Package cleanly when asked.
For GitHub or sharing, keep one canonical skill folder with:
SKILL.md
agents/openai.yaml when useful
references/ only for material worth loading on demand
scripts/ only for deterministic helpers
Codex-Native Polish Checklist
Always review installed Claude skills for these issues before calling the port done:
- Frontmatter uses simple YAML with a clear one-line
description unless parser validation proves folded text works.
- Description names realistic Codex trigger phrases, not only Claude slash commands.
- Body says Codex where the agent is Codex; keep "Claude" only when it names an external tool or upstream source.
- Remove
.claude-plugin, marketplace, hook, SessionStart, and UserPromptSubmit expectations from the installed skill.
- Rewrite persistent behavior claims such as "ACTIVE EVERY RESPONSE" so they respect Codex skill triggering and conversation scope.
- Add Codex priority rules when a skill changes assistant behavior: system, developer, safety, and tool instructions win.
- Preserve exact text requirements for code, commands, file paths, URLs, errors, test output, and citations.
- Convert host-specific examples such as bare
python3 to commands that fit the user's OS when installing locally.
- If helper scripts call Claude, Anthropic, OpenAI, or another external backend, document that dependency and the data boundary honestly.
- Remove broken mojibake or non-ASCII examples unless they are necessary and verified.
- Install one canonical skill folder; do not copy whole Claude plugin repos into Codex when wrappers add no value.
Fast Path For GitHub Skill Repos
When the user provides a GitHub repo or skill URL:
- Inspect the repo root and
skills/ first.
- Also inspect
.claude/skills, root skill.json, and .claude-plugin/plugin.json before deciding the repo is only docs or a catalog.
- If the repo, owner, path, or surrounding text carries Claude-oriented signals such as
anthropics/skills, Claude, Anthropic, or .claude/*, stay in this toolkit flow even when the skill path looks obvious.
- Install the exact skill folder, not the whole repository, when the target path is clear.
- Read the installed
SKILL.md immediately after install.
- Run the Codex-native polish checklist even if discovery is already clean and trigger quality is acceptable.
- Switch to migration if the installed skill is invisible, vaguely named, malformed, weakly described, or still Claude-only.
Bias toward exact --repo plus --path installs when you can infer the path confidently.
Fast Path For Local Claude Shell Snippets
When the user provides a shell snippet like mkdir ... .claude/skills/... && printf ... > skill.md:
- Treat it as a local scaffold, not proof that a published skill was installed.
- Inspect the generated folder directly.
- Normalize
skill.md or Skill.md to SKILL.md.
- Convert markdown metadata sections into YAML frontmatter when needed.
- Rewrite any Claude-specific persistence, hook, plugin, or command assumptions.
- Only after normalization and Codex-native polish should you claim the skill is Codex-ready.
Migration Rules
When migration is required:
- Preserve behavior, not original packaging.
- Rewrite the frontmatter so
name and description clearly express both function and trigger phrases.
- Add Codex priority rules when a skill changes assistant behavior, especially for safety, exact output, code, tests, and tool results.
- Replace Claude-only commands with Codex-appropriate actions or mark them as external dependencies.
- Keep the body concise and operational.
- Move long playbooks, mappings, and troubleshooting into
references/.
- Avoid duplicate aliases unless there is a real discovery reason.
For the full sequence, read references/migration-playbook.md.
Discovery Rules
- Treat "installed on disk but missing from
codex debug prompt-input" as a failed install.
- Save
SKILL.md as UTF-8 without BOM.
- Prefer one canonical folder per skill.
- Keep frontmatter simple.
- Make the
description trigger-rich enough that Codex can match real user phrasing.
If discovery is broken, use references/discovery-troubleshooting.md.
Packaging Rules
When building a shareable Codex-native port:
- publish the canonical skill folder itself
- avoid extra documentation files unless the user explicitly wants repo-facing docs
- keep helper scripts generic and deterministic
- do not ship duplicate migrated variants of the same skill
- if a repo contains multiple skills, package each one as a distinct skill folder
Required Validation
Before calling the migration finished, verify:
scripts/check_skill_md.py <path-to-skill-md> reports no blocking issues
scripts/inspect_skill_repo.py <path-to-skill-folder> shows the expected structure
- compatibility score, dependency profile, and install tier are sensible for the chosen route
- trigger quality is good enough for realistic prompts, or the report clearly explains why migration polish is still needed
codex debug prompt-input shows the skill when installed into a live Codex skills directory
Reference Guide
Read only what is needed:
references/migration-playbook.md for the end-to-end migration sequence
references/compatibility-matrix.md for route selection
references/install-route-adaptation.md for translating Claude install tutorials into Codex actions
references/install-workflow.md for the semi-automatic installer pattern
references/discovery-troubleshooting.md for failure analysis