16.7 One skill, three bundles
- Approved by
- Panaversity ·
In everyday life. A recipe written in plain English works in any kitchen. The way you deliver a meal kit does not. One grocer wants a box with its own label, another wants a bag with a barcode, and a third wants you to bring the ingredients yourself. The recipe card goes in all three unchanged. Only the packaging differs.
What you will learn. What moves from one AI tool to another and what does not. How to keep one copy of each skill for all of them. How to test it in every place you promise.
Why it matters. A team rarely uses one AI tool. If each tool needs its own copy of a skill, the copies drift, which is Grace's problem from the overview again.
Which three words do I need?
- A host is the product that loads a skill or a plugin.
- A surface is a place where a host runs: a chat in a browser, a task on your desktop, a coding agent in a folder.
- A building agent is the coding agent that writes the files.
A plugin is only files, so any building agent can write any host's files. You can build a plugin for one company's products with another company's coding agent.
What moves, and what does not?
| Part | Moves? | Why |
|---|---|---|
| A skill | Yes, to every host in this chapter | All of them follow the Agent Skills standard (16.1) |
| The plugin's label and layout | Only to hosts that follow Agent Plugins | Some hosts read only a manifest of their own |
| A connector that runs on your computer | Only to surfaces that can start a program there | A chat in a browser cannot run a program on your laptop |
| A hook | Least of all | Each host has its own format |
How do I serve many hosts with one plugin?
Keep one copy of each skill, and add each host's own label beside it. Never copy a skill into a second folder, because two copies drift. A connector can also serve the same skill to any AI app that connects, with a list of its files that the AI app checks (Chapter 20). It is still one copy, served from one place.
plugin/
├── plugin.json the standard label, for hosts that read it
├── mcp.json
├── skills/ one copy, shared by every host
└── .host-a/plugin.json one host's own label, pointing at the same skills/A host that reads no plugin at all, only a skills folder named in its settings, is served too: its settings point at the same skills/. If a host starts to follow the standard later, you delete its own label, and nothing else changes. Your coding agent writes each host's label, and the checks from 16.6 prove each one is well formed.
Does a skill behave the same everywhere?
No, because surfaces differ. A coding agent in a folder has your repository. A chat in a browser does not. So a skill that says "run the repository's triage command" works in the first and fails in the second.
That is why Brightline's triage skill carries a small script of its own in scripts/. A test proves the script gives the same answer as the repository's tested code, and the skill names both routes.
How far do I test?
In every place you promise. Send the same request in each one, and write down three things, kept apart: did the skill load, did it run, and did its answer match the script.
At work. Brightline wants the triage skill in all six places its staff use AI: three surfaces from one company, two from another, and an open-source coding agent. The builder writes the skills once, and they ship as three bundles. The first company's hosts read their own manifest, beside the standard one. The second company's read the standard one. The open-source agent's settings point at the shared skills folder. The data connector works in the coding agents and the desktop task. In the browser chats it is skipped, and people attach the files instead. The skill handles both: if the connector is there, it reads the folder, and if not, it asks for the files.
Try it
You take the vendor-terms skill to a second place. You need the plugin from 16.5.
Make. On paper, draw this table, with one row for each place you try:
| Place | Loaded | Ran | Answer matched |
|---|
Predict. Choose a place you have not used yet from the tabs in 16.10, and follow its steps. If you have only one agent, use the Claude or ChatGPT tab, in a browser or the desktop app. Before you send anything, decide which of the three columns this place will fill.
Run. In a new chat or session, send the same request as before: "What terms do we have with Great Lakes Freight?" Fill in the row.
What you should see. In a coding agent in the folder, all three, and Net 30. In a chat with no repository, the skill loads, but npm run vendors has nothing to run in. A skill that loaded but could not run has moved only part of the way. Write it down as that.
Modify. Now do what Brightline did. In practice-16, send: "Give the vendor-terms skill a script of its own, at practice-plugin/skills/vendor-terms/scripts/vendors.py, with a copy of the vendor list beside it and no dependencies. It prints the same lines as npm run vendors. Change the skill's body to name both routes: npm run vendors in this repository, and the script anywhere else. Then run both, and show me that they print the same lines."
Take the plugin to the second place again, and send the same request. The row should now fill all three columns.
The operating rule: write one copy of each skill, and put each host's own label beside it. Never copy a skill. Test the same request in every place you promise.

Figure 16.11. Any building agent can write any host's files. The skills travel everywhere. The bundle around them is read in different ways, and each surface loads a different part.
Building Skills and Plugins with Coding Agents
Question 1 / 8
0 answered
Can one company's coding agent build a plugin for another company's product?
- Owner
- Panaversity