SkillBridge is a working MVP Codex skill for converting supported Claude Skill folders into Codex-ready local skill folders.
It reads a Claude Skill, classifies its execution model, preserves detectable functionality, copies safe resources, generates Codex-native handoff docs, and runs safe smoke tests when possible. It is intentionally conservative: it does not install dependencies, run package managers, create plugin files, or claim full automation readiness from partial evidence.
Use this project if you:
- Have Claude Skills you want to run as Codex Skills.
- Need a repeatable conversion workflow instead of a hand rewrite.
- Want generated Codex skills to preserve scripts, references, assets, templates, and licenses where safe.
- Want clear labels for workflows that are tested, untested, dependency-blocked, or user-input-blocked.
SkillBridge is a working MVP for the supported Claude Skill patterns validated so far.
Manual Codex validation has passed on four representative converted skills:
| Skill | Source type | Converted skill behavior | Manual result |
|---|---|---|---|
| brand-guidelines | instruction-heavy | Applied brand rules to create a landing hero | PASS |
| theme-factory | asset/template-heavy | Used copied theme showcase and theme files to create a custom theme | PASS |
| frontend-design | frontend/design-output | Created and checked a standalone HTML/CSS hero | PASS |
| script-backed/file-processing | Ran dependency-free bounding-box validation | PASS |
See docs/manual-validation-results.md.
- Instruction-only and instruction-heavy skills.
- File-generating skills.
- Script-backed local workflows.
- Asset/template-heavy skills.
- Hybrid skills with mixed instructions, scripts, references, and assets.
- Dependency detection and dependency setup notes.
- Safe fixture generation for local smoke tests.
- Safe smoke execution when the workflow is local, dependency-safe, path-safe, and fixture-ready.
- Translation-first workflow preservation with dropped workflow accounting.
- Not every converted workflow is automatically tested.
- Dependencies are not installed automatically.
- Package managers are not run.
- Network, auth, browser, MCP, paid-service, and credentialed workflows are not proven during conversion tests.
- Subjective creative quality still needs human review.
- API-backed workflows require user setup.
- Plugin packaging is not included yet.
Clone this repo, open Codex in the repo, and invoke $skillbridge. Point it at a Claude Skill folder and choose an output folder for the converted Codex skill.
For direct CLI use from the repo root:
python .agents/skills/skillbridge/scripts/convert_skill.py `
--source "C:\path\to\claude-skill" `
--output "test-outputs/codex-skills/my-converted-skill" `
--overwriteValidate the generated Codex skill:
python .agents/skills/skillbridge/scripts/validate_skill.py `
--skill-dir "test-outputs/codex-skills/my-converted-skill"Inspect:
SKILL.mdreadiness-report.mdreferences/functionality-preservation-map.mdreferences/validation-plan.md
Install locally by copying the converted skill folder into one of:
<repo>/.agents/skills/<skill-name>
C:\Users\<user>\.agents\skills\<skill-name>
Then start a fresh Codex instance and explicitly invoke the converted skill.
A converted skill typically contains:
<skill-name>/
SKILL.md
readiness-report.md
scripts/
assets/
fixtures/
outputs/
references/
dependency-notes.md
functionality-preservation-map.md
functionality-preservation-map.json
output-contract.md
validation-plan.md
Only relevant folders are generated or populated.
The converter:
- Does not install dependencies.
- Does not run package managers.
- Does not create plugin files.
- Does not run network/auth/browser/MCP/paid-service workflows.
- Keeps outputs under safe generated folders.
- Separates translation coverage from smoke-test evidence.
- Keeps task readiness and automation readiness conservative unless explicitly proven.
Lightweight example prompts and source summaries live under examples/.
Generated conversion folders are not committed by default. They are intentionally ignored because they can be large and local-machine-specific.
- Test more third-party Claude Skills.
- Add optional metadata such as
agents/openai.yaml. - Add optional plugin packaging in a separate release step.
- Improve generic fixture generation for dependency-heavy PDF/API workflows.
- Add broader external-service/API conversion examples with safe setup notes.
- Improve subjective quality evaluation workflows for creative skills.
MIT. See LICENSE.