Agent Skills portability after the AAIF donation: test one bounded task first

The October 22 change concerns stewardship, not identical host behavior

The original Anthropic announcement donates Agent Skills to the Linux Foundation’s Agentic AI Foundation and says using or contributing does not require joining the foundation. Versioned releases, a proposal process and security guidance are described as a proposed roadmap, rather than features already implemented everywhere.

The Agent Skills specification defines the SKILL.md directory format and required name and description. Optional compatibility describes environment requirements; allowed-tools remains experimental. A declaration in a file cannot itself grant host permissions or demonstrate an executable capability.

Prepare one delivery-gap task

This is an original editorial reference exercise, not a comparison run on two products. Use invented records you are entitled to supply. It needs only a returned note, letting the first test focus on information rather than real email, repositories or customer material.

R1: the ZIP opens; two of three required screenshots were found; owner Lin; next check is to identify the missing screenshot. R2: the workbook contains two worksheets while the handoff requires three; the record says Friday 16:00 without a date or time zone and supplies no owner. R3: all three attachments were found and opened, with nothing unresolved.

Start with a small fragment and explicit boundaries

The following original suggestion lets readers create their own specimen. It is not an installed resource or an official product skill. Match the folder name to name and follow the current naming constraints. No script or external dependency is included, making initial failures easier to locate.

---
name: handoff-gap-note
description: Turn supplied delivery checks into a short gap note. Use when asked to summarize missing handoff evidence.
---
Use only the supplied records. For each unresolved record, output ID, observed, expected, owner, and next check.
Preserve IDs. Mark missing fields as unknown. Do not invent dates or turn an observation into a completed fix.
If a required reference file is absent, name the missing file and stop that part of the task.
Do not run commands, install packages, send messages, or alter files. Return the note for review.

Three cases expose different failures

Give each host the same R1–R3 records. Do not supply raw records to one and the first generated summary to the other. Check that only R1 and R2 appear, IDs and numbers survive, Lin remains with R1, and Friday remains incomplete. Fluent formatting cannot compensate for omitted facts.

Next, deliberately add a requirement to read references/checks.md without supplying that file. The reference behavior is to identify the missing dependency and stop that part, rather than inventing a checklist. Finally request an unrelated birthday greeting and inspect whether the gap-note workflow activates incorrectly. Discovery and invocation differ by host; follow that host’s documentation.

Do not redefine success to hide a difficult case

An output saying “Lin fixed R2 before Friday 16:00” makes three unsupported moves: it transfers R1’s owner, resolves an incomplete time and changes an open issue into completed work. Repair the evidence rules and unknown fields, then rerun rather than deleting R2.

A complete plugin has separate installation, dependency and submission questions; see Claude plugin submission readiness. For repeatable task material, see the meeting-action skill input and template. Neither replaces acceptance of this task using the same evidence.

Keep migration evidence rather than a compatibility slogan

Record host and client versions, model, skill version, input, output and failed cases. Adding scripts or network dependencies requires separate environment, execution-permission and outcome checks. A script-free test does not establish that an executable workflow also passes.

The reference here was derived manually from the supplied records. No two-product test, time saving or universal account access is claimed. Define acceptance through the task: required information survives, missing material is not silently invented, and actions outside the task retain their real authorization requirements.