Are Agent Skills actually portable? — five-client evidence map #10

Five major AI coding clients now document some form of Agent Skills support. That does not prove that one skill behaves the same across them.

The new Doubt evidence map separates the portable core from the client-specific behavior:

• 5 claims
• 6 evidence items
• 6 official sources
• 1 contradiction
• 1 explicit unknown

The practical verdict: author one conservative core, then install and test it per host. SKILL.md, references, assets, and scripts may be portable; discovery, activation, consent, tool use, and output behavior may not be.

Explore the evidence:

• Interactive map
• Machine-readable .doubt.json
• Receipt: 89f18aa3a3b2b9136096f772b3a884b89f2529e9dae31f4aab721ed190fc4908

The map also exposed the next missing edge: no same-fixture behavioral benchmark across Claude Code, Codex, GitHub Copilot, Cursor, and Gemini CLI.

If you can run even one client, contribute a reproducible result to benchmark issue #9. Positive, neutral, and negative evidence are all useful.