Portable AI skills still need an owner

Teams often repeat more than the work itself.

They repeat the prompt, attach the same reference documents, explain the output format and remind an AI tool which checks matter. The process may work for one person in one chat, but the useful knowledge remains difficult to share, review or improve.

Agent Skills offer a more durable shape for that knowledge. A skill packages instructions and supporting resources into a reusable playbook that a compatible AI agent can load when the task needs it.

That can make one stable process easier to repeat across a team and, increasingly, across AI products. It does not make the process reliable, secure or portable by itself.

The business still needs to own the workflow behind the file.

What changed on 22 October 2026

On 22 October, the Linux Foundation announced that Agent Skills had joined the Agentic AI Foundation. Anthropic contributed the project to the foundation, where a multi-company technical steering committee will guide the open standard.

The format is intentionally simple. The current Agent Skills specification defines a folder with a SKILL.md file containing metadata and instructions. A skill can also include scripts, reference material, templates and other assets. Compatible agents discover the skill from its name and description, then load the detailed instructions only when a task calls for them.

The format is already moving beyond one vendor.

OpenAI's guidance for using skills in ChatGPT describes SKILL.md as a portable, versionable playbook for recurring work. Google announced skills for Gemini and Workspace on 30 September, then started the Gemini app rollout on 13 October. Google says its skills use the same open SKILL.md standard.

The Google rollout also shows why the details still matter. Skills in the Gemini app do not currently synchronise with skills in Workspace apps. The Workspace Studio and Workspace app launch is delayed, while the Gemini app rollout is gradual and expected to complete by mid-November. Google also plans to retire Gems for business and enterprise users no sooner than 1 March 2027, with further notice before the change.

The useful development is not another prompt feature. It is the emergence of a shared format for business instructions, alongside real differences in availability and behaviour between the products that use it.

Why this matters to a smaller business

Small teams carry important operating knowledge in people's memory, old documents and copied prompts.

That can include how to:

• triage a new enquiry and prepare the right context for review;
• turn meeting notes into a draft action register;
• check website content against brand, accessibility and search requirements;
• assemble a recurring operations or performance report;
• compare a supplier against agreed technical and commercial criteria;
• prepare a customer reply from approved source information.

A skill can make the process visible in plain text. People can review the steps, improve the examples and keep the same quality checks when another colleague or compatible AI tool performs the work.

This creates a credible AI Workflow Development opportunity when the process already happens often and the team agrees on what a useful result looks like.

It is less useful when the process changes every week, depends mainly on undocumented judgement or has no clear source of truth. In those cases, writing a skill can formalise confusion instead of removing it.

Treat the skill as an operating procedure

A good first skill should describe work the business already understands.

Before writing instructions, record:

• the purpose of the task and the person accountable for it;
• the inputs the workflow needs and their approved sources;
• the sequence people follow now;
• the decisions AI may assist and the decisions a person must make;
• the expected output, including its format and quality checks;
• the exceptions that should stop the workflow and ask for help;
• the current time, correction effort, error rate or response delay.

This is the operating procedure. The SKILL.md file is one implementation of it.

That distinction matters. If the business process exists only inside instructions written for one AI product, a future migration still requires the team to rediscover why each step exists.

AI Business Automation should clarify and simplify the workflow before packaging it for repeated use. The aim is not to make every task an AI task. It is to remove repeated explanation from work that already has a stable purpose.

Portable instructions are not a portable workflow

An open file format reduces one kind of lock-in. It does not remove the rest.

A skill can carry instructions, examples, references, templates and scripts. It does not automatically carry:

• the identity used to access a business system;
• the permissions granted by each AI product or connected app;
• the provider's retention, training and regional processing terms;
• the same model behaviour or tool selection;
• a hosted search, browser or document tool with equivalent limits;
• conversation history, approval records or audit logs;
• an assurance that the skill activates at the same time in every product.

The specification makes some of this variation explicit. Its compatibility field can describe product, software or network requirements. Its allowed-tools field is experimental, and support may vary between implementations.

Google's current lack of synchronisation between Gemini and Workspace skills is a practical example. The file may be portable while deployment, sharing and administration remain product specific.

Use portability as a testable property, not a marketing assumption.

Run the same skill against representative work in each intended product. Compare whether it activates correctly, follows the important steps, uses the expected sources, respects the review point and produces an output the next system can accept. Record any product specific instructions or tools separately so the common business procedure stays readable.

Keep the skill small enough to understand

A skill should not become a hidden application written in prose.

Start with one bounded result, such as a draft enquiry summary or a content quality checklist. State the required inputs, the steps, the output format and the final checks. Add examples only where they resolve a real ambiguity.

OpenAI's guidance recommends small building blocks instead of one large end-to-end skill. The Agent Skills specification uses progressive disclosure for the same reason: the agent begins with a short description, loads the main instructions when needed and reads supporting resources only when the task requires them.

Smaller skills are easier to:

• review with the person who understands the work;
• test against expected results and failure cases;
• reuse without exposing unrelated information;
• change without affecting several processes at once;
• retire when the workflow no longer exists.

If the task needs complex state, permissions, records, approvals or an interface, a skill may be only one layer. The right container may be an AI Integration or a focused internal tool rather than a longer instruction file.

Review skills before installing them

A skill is readable, but that does not make every skill trustworthy.

The specification allows a skill to include executable scripts in languages such as JavaScript, Python or shell. The Markdown body can contain any instructions that help the agent perform the task. That flexibility is useful, and it means a third party skill should be reviewed as both instructions and software before it is allowed near business information or systems.

Check:

• who published it and whether the source is still maintained;
• every instruction the agent will load;
• bundled scripts, dependencies and external network destinations;
• files, credentials and environment information it expects to access;
• the tools and permissions it requests;
• what it writes, sends, deletes or changes;
• how updates are reviewed before reaching the team.

Test an unfamiliar skill with non-sensitive data and limited access. Do not place passwords, API keys, customer records or confidential material inside the skill folder.

The Australian Signals Directorate's small business guidance for using AI securely recommends reviewing which systems, accounts and data an AI agent can access, granting only what it needs and retaining human oversight for significant changes. The same control pattern applies when a skill tells the agent how to use those capabilities.

The existing guide to technical boundaries for AI agent pilots explains why the identity, permissions and available tools must enforce the limit. Instructions can describe the boundary. They should not be the only thing holding it in place.

Give each business skill an owner

Shared instructions become stale like any other operational asset.

Assign one owner who can answer:

• Which process does this skill represent?
• Which version is approved for team use?
• Which products and models has it been tested with?
• Which source documents are current?
• Who reviews changes to instructions or scripts?
• What evidence shows the output is still useful?
• When should the skill be reviewed or retired?

Keep the skill in a versioned location where changes can be reviewed. Record the reason for material updates. Maintain a small set of representative cases that includes normal work, incomplete information, conflicting sources and situations that must reach a person.

Model and platform changes can still affect the result even when the skill file does not change. The production AI model change guide describes how representative tests, a limited rollout and a fallback make those changes easier to manage.

This does not need a large governance programme. It needs enough ownership to prevent an old instruction from quietly becoming the team's current process.

Keep privacy and human judgement outside the prompt

A skill should state which information is necessary, but the surrounding system must control whether that information is available.

Before connecting documents, inboxes, customer records or internal systems, confirm:

• why the information is needed for this task;
• whether a less sensitive source would be sufficient;
• where the selected product processes and retains it;
• who can install, share, edit and run the skill;
• which outputs are stored and who can see them;
• how access is removed when a person or workflow changes.

The Office of the Australian Information Commissioner recommends privacy by design, due diligence, testing, human oversight and ongoing monitoring for commercially available AI products.

Human judgement should remain where an output changes customer expectations, employment or financial records, access, legal rights, public content or another consequential business decision. A reusable skill can make the preparation more consistent. It does not become accountable for the outcome.

A practical first skill

For one stable, repetitive process:

1. Record the current time, correction effort, error rate or response delay.
2. Define the approved inputs, expected output and person accountable for the task.
3. Write the smallest skill that captures the steps, examples and final checks.
4. Connect only the sources and tools the task genuinely needs.
5. Begin with a read-only or draft role and keep human approval before consequential actions.
6. Test representative work, edge cases, missing information and unsafe requests.
7. Try the skill in each intended product and document what is not portable.
8. Version the approved skill, measure the complete workflow and improve it before expanding use.

If the team cannot agree on the procedure or recognise a useful result, the process is not ready to become a shared skill.

The next decision

Agent Skills make it easier to carry a business playbook between people and compatible AI tools. The open standard is useful because the instructions can remain visible, versioned and less dependent on one prompt editor.

The business value still comes from the process behind the file.

Start with one repeated task that the team understands. Give it an owner, approved sources, representative tests and a clear human decision point. Prove that the skill reduces repeated work without hiding responsibility or creating a new access path.

An AI Workflow Development engagement can turn one stable process into a focused, testable skill. When the decision spans several tools, teams or business systems, a Fractional Technical Partner can help define what should be portable, what must remain controlled and how the workflow will be maintained after the first useful result.