Compatibility Checklist

Use this checklist when auditing a Claude skill for ChatGPT/Codex translation.

Source identity

• Does the directory name match the frontmatter name?
• Is there exactly one SKILL.md entrypoint?
• Does the description explain real trigger conditions?
• Is the skill broad enough to collide with other skills?
• Does the skill contain host-specific statements that should remain only because Claude is the subject of analysis?

Target host

Identify the intended target before translating:

• chatgpt: package for ChatGPT skill use, with ChatGPT UI metadata.
• codex: package for coding-agent workflows, repository context, shell/script use, and engineering constraints.
• both: preserve the portable core and call out host-specific execution notes.

Do not assume ChatGPT and Codex have identical runtime capabilities.

Host assumptions

Flag any mention of:

• Claude Desktop, Claude Code, Anthropic API, Projects, Artifacts, or Claude-only UI behavior.
• MCP servers or tools that may not exist in the target host.
• local filesystem paths, desktop apps, shell commands, package managers, OS-specific commands, browser profiles, or background processes.
• connectors such as Google Drive, Notion, Slack, GitHub, Linear, Dropbox, SharePoint, or email.
• credentials, API keys, OAuth state, cookies, logged-in sessions, or vendor accounts.
• background/asynchronous work promises.
• platform-specific install paths, marketplace behavior, or package formats.

Resource portability

For every file:

• Keep text references that are still relevant.
• Keep scripts only when the target host can run them or Codex can execute them.
• Keep assets only when they are needed for output.
• Remove placeholder files.
• Avoid deeply nested reference chains.
• Check package size risk before packaging.

Capability classification

Use these definitions:

• portable: behavior and resources move unchanged.
• adaptable: equivalent behavior is possible after rewriting metadata, paths, host names, or instructions.
• host-dependent: possible only if the target has a matching tool, connector, runtime, or credential.
• unrealizable: cannot be delivered truthfully in the target environment.

Risk score

Assign one overall score:

• low: all material capabilities are portable or trivial metadata adaptations.
• medium: behavior is preserved, but instructions, resource organization, or target notes need meaningful adaptation.
• high: material capabilities require target-specific tools, connectors, runtime setup, credentials, or shell behavior.
• blocked: an essential capability cannot be preserved and no reduced version should be produced without explicit user approval.

Capability matrix

Always produce a matrix before translation:

| Capability | Claude/source assumption | ChatGPT/Codex equivalent | Status | Notes |
|---|---|---|---|---|
| Summarize uploaded CSV | Claude can read attached file | ChatGPT can read uploaded file if provided | portable | no change |
| Query local SQLite DB | Claude Desktop MCP server configured | No equivalent unless user configures MCP/tool | host-dependent | approval required |

Approval gate

Ask the user before packaging if:

• any capability is host-dependent.
• any capability is unrealizable.
• the risk score is high or blocked.
• the translation must become a minimum viable translation.

Use direct wording:

I can translate this skill, but these capabilities will not be equivalent in ChatGPT/Codex:
- ...

Proceed with the adapted or minimum viable version?

ChatGPT/Codex target requirements

The translated skill should include:

• SKILL.md
• agents/openai.yaml
• references/ only when useful
• scripts/ only when useful and portable
• no unused placeholder files

The frontmatter must contain only:

---
name: example-name
description: lower-case trigger-focused description
---

Anti-overbroadening rule

Do not broaden the translated skill's purpose to compensate for missing tools.

Bad:

If Notion is unavailable, use any available source to answer.

Better:

This Notion-dependent capability requires Notion access in the target host. If unavailable, state the limitation and ask the user to provide the needed content directly.

Round-trip review

After packaging, report:

• What did the original skill do better?
• What became weaker?
• What target-host assumptions remain?
• Which test prompt should verify equivalent behavior?