A business often gets its best AI results from knowledge that is not in the model: the way it checks a customer brief, structures a proposal, or labels a finished asset. Keeping that knowledge in a long chat makes it difficult to maintain. Copying it into five assistants creates five versions that can quietly diverge.
Portable agent skills offer a different approach: package a procedure with the references and resources needed to carry it out. The practical trend is not that prompting has disappeared. It is that repeatable prompting is becoming a maintained work asset.
What changed in 2026
On October 22, the Linux Foundation announced that Agent Skills, contributed by Anthropic, was joining the Agentic AI Foundation. The announcement describes an open format for reusable instructions, scripts, and resources that agents can load on demand.
The Agent Skills specification gives the idea a concrete shape. A skill is a directory containing a required SKILL.md file, with metadata and instructions. It can also contain scripts, reference documents, and assets. The specification describes progressive disclosure: an agent can initially see identifying information and load fuller instructions when the task calls for them.
This is a useful separation. A general assistant need not keep every company procedure in its active context all the time. It can discover an appropriate procedure and then retrieve the details.
Start with a process you can explain
Good candidates have a recognizable trigger, a small set of inputs, and an inspectable output. Examples include turning an approved interview transcript into a case-study draft or checking whether a product description follows a house style.
Bad starting points are instructions such as “handle marketing” or “make our business successful.” Those leave success, authority, and stopping conditions undefined. A skill cannot repair a process whose owner cannot say what a correct result looks like.
A useful package should state required inputs, reference priorities, output format, and situations requiring human review. Keep changeable facts, such as a product catalog, separate from relatively stable writing instructions.
Example: a reusable case-study draft
Consider an agency that wants consistent customer stories. This is an illustrative design, not a workflow we tested.
Its skill could contain a structure for the article, a tone guide, and a checklist for attributing claims. The instructions would require an approved transcript and prohibit inventing performance figures. Missing outcomes would become questions for the editor, not plausible-looking statistics.
The output might include a draft, a list of claims with transcript references, and unresolved permissions. Publishing would remain a separate, explicitly approved step. The skill therefore packages editorial judgment without silently gaining authority to contact the customer or post to a website.
Portability is not automatic execution
An open format helps transfer instructions, but an agent still needs the tools and environment those instructions assume. A script requiring a particular runtime will not necessarily work in a browser-only assistant. A reference to a company drive is useless without appropriate access.
The specification includes an optional compatibility field and an experimental allowed-tools field. Buyers should not interpret a tools declaration as a universal enforcement mechanism. Actual permissions depend on the host product.
Treat downloaded skills like other third-party software. Read their instructions, inspect executable files, check licenses, and avoid packages that request broad access unrelated to the job. A friendly description is not a security review.
Measure improvement before sharing widely
Anthropic’s January evaluation guide recommends clear success criteria and a mix of deterministic checks, model grading, and human judgment. Apply that thinking to skills: compare results with and without the package on representative tasks.
For the agency, check factual attribution, structure, editing time, and appropriate refusal when the transcript lacks permission. Include a case where the skill should not activate. An eager skill that hijacks unrelated work is not a successful one.
Watch next for consistent behavior across host products and clearer dependency handling. The most valuable skills will not be the longest instruction files. They will be the ones a team can understand, version, evaluate, and safely reuse when its preferred AI tool changes.
Sources
- Linux Foundation: four new AAIF projects — 2026-10-22; accessed November 7, 2026.
- Agent Skills specification — undated documentation; accessed November 7, 2026.
- Anthropic: Demystifying evals for AI agents — 2026-01-09; accessed November 7, 2026.
