Your AI Skill Now Works in Every AI Tool. Here's How to Build One.

You have been building on rented land.

Priya Anand runs a five-person bookkeeping and tax-prep practice in Manchester. Every new client goes through the same ritual: proof of ID, an engagement letter, prior-year returns, a folder set up in the shared drive, three follow-up emails if anything is missing. It took her four years to get that checklist into her own head. It took her one afternoon to teach it to Claude as a Skill.

Then her firm's ops lead, who prefers ChatGPT, asked to use it. Priya could not hand it over. The instructions she had written lived inside one app. If the firm ever changed AI tools, or if a new hire preferred a different one, she would have been rewriting the same checklist from scratch, again.

As of this month, that is no longer true for a new class of AI automation. Here is what changed, and how to build something that survives whichever tool your business, or your team, is using next year.

The deal you didn't know you signed

Every AI skill built before August 2026 came with a hidden clause. You taught the AI your business, one instruction at a time, and in exchange you agreed to stay. The knowledge did not travel. Ask a search engine today whether an automation survives a tool switch and the honest answer, echoed across Google's own AI summary, is blunt: "you must rebuild your workflows from scratch on the new platform."

That has been true for years, and most small business owners have simply accepted it as the cost of automating anything. Write the prompt, live with the platform.

What actually shipped on August 6

On August 6, 2026, a new open specification called Agent Plugins reached version 1.0.0. It is not a product. It is a folder format, and it is short enough to read in five minutes at agent-plugins.org: one manifest file called plugin.json, one optional folder of skills written as plain-language SKILL.md files, one optional mcp.json file for any live tool connections. That is the entire specification.

One widely shared post announcing the launch put it plainly: "Agent Plugins: Build it once, use it everywhere." It pulled over 82,000 views and 624 likes in a day, according to its public metrics on X, and by the next morning a separate post was counting 1.1 million views for what one developer called, correctly, "a file format." Within a week, GitHub Copilot, VS Code, ChatGPT, Codex, Cursor, and Amazon's Kiro all supported it, and independent builders reported Hermes and OpenClaw joining the list.

What is the Agent Plugins standard, in plain terms?

Agent Plugins is a shared packaging format for two things AI companies had already invented separately: Agent Skills, the plain-language instruction files that tell an AI how you want a task done, and MCP, the protocol that lets an AI actually connect to a live tool, like your calendar or your accounting software. Before August 6, every AI client stored those two things in its own private folder layout. A skill built for one tool needed rebuilding, sometimes from a blank page, before another tool could read it.

Agent Plugins does not touch how any single tool works. It only standardises the folder around the work you already did, the way a shipping container did not change what factories made, only how what they made got carried.

Isn't this just Claude's homework, standardised without Claude?

Fair question, and worth asking before you act on any of this. Claude was not one of the five founding members, and more than one person online pointed that out within hours of the launch. But look at what the standard actually copied: a manifest file, a skills folder, plain-language SKILL.md instructions. That layout is, almost line for line, the plugin structure Claude Code already used. A demonstration built the week of launch installed the identical folder into both Claude Code and OpenAI's Codex with zero edits between them. In practice, a skill written for Claude today is already close to portable; Agent Plugins mostly formalises what was already true and extends it to five more tools.

A second fair question: is it safe to install a plugin someone else built? One developer's caution from the same week is worth repeating honestly: back in June, before this standard existed, a fake Agent Skill reportedly passed every automated security scanner and reached thousands of installs before it was caught. Agent Plugins standardises the folder, nothing more. It does not vet who built what is inside it. Treat an unfamiliar plugin the way you would treat a new hire's laptop: know who built it, read the skill file before you install it, and prefer plugins from people or companies you can actually name.

How to build one skill your whole team can use

You do not need to be a developer for this. You need the same thing Priya already had: a task you or your team repeats, written down in plain English.

1. Write the checklist first. Not the technology, the actual steps. What Priya wrote for client onboarding: request ID and engagement letter, request prior three years of returns, create the client folder, send a reminder after 48 hours if anything is missing.
2. Save it as a Skill. Make a folder named for the task. Inside it, one file called SKILL.md holding your plain-language instructions. This is the exact format Claude Skills already use, so if you have written a Claude Skill before, you have already done this step.
3. Add the plugin wrapper. One more file, plugin.json, in the same top-level folder, with a name and a one-line description. The specification caps it at ten possible fields. That single file is what turns a Skill folder into a portable Agent Plugin.
4. Install it in whatever tool you use. In Claude Code: /plugin marketplace add yourname/your-repo, then /plugin install your-skill@yourname. In OpenAI's Codex, the same two steps, with codex in place of /plugin.
5. Hand the identical folder to a teammate on a different tool. No edits, no translation, no second version to maintain. That is the entire pitch of Agent Plugins, proven on camera the week it launched.

What you have after this

A week from now, you have one real Skill, written down properly for the first time, sitting in a folder instead of scattered across your own memory and three people's inboxes.

A month from now, that folder is installed in whatever tool your team actually prefers, not the one tool you happened to pick first, with nobody arguing about which AI product the business should standardise on.

Two quarters from now, when a sharper model or a cheaper AI subscription shows up, and one will, moving to it is an install command, not a rebuild project.

Same knowledge. Any app. Nobody has to agree on a tool anymore.