Picture the capabilities your teams rely on outliving every platform decision you make. Finance runs its Monday receivables summary on the AI platform it has today. Next year the company standardizes on another, or runs two side by side, and the same capability is in service there within days, to the same standard, with the same tests proving it. Your organization’s know-how compounds year after year, on whichever of ChatGPT, Claude, Copilot or Gemini your people open in the morning.
That is the plan to make now, because the platform question never stays settled. Vendors ship, contracts renew, an acquisition arrives with a different workspace. The CTO who can say exactly what moves and what gets adapted keeps control of how the work gets done.
Every capability has two layers.
The first layer is the method: what the work is, which sources count, the rules your best people apply, what a finished output contains and how a reviewer judges it. It is written in plain language and belongs to your organization. What is an AI capability? sets out its parts.
The second layer is the harness: how the method reaches your systems and your people on one particular platform. It covers the connections to data, the permissions behind them, what starts a run, how the instructions are packaged and shared, and the test record on that platform.
The method carries over as written. The harness is adapted, platform by platform: a short, known piece of engineering once you have inventoried it.
The inventory for one capability.
An invented 90-person architecture firm built a weekly receivables summary on the platform its finance team already used. The firm is now standardizing on a second platform. Its IT lead and controller wrote this inventory before moving anything.
Illustrative example with invented data. The weekly receivables summary at an invented architecture firm, inventoried before a move to a second platform.
| Carries over as written | Adapted per platform |
|---|---|
| Instructions. Read the aging report export and the collections notes. Group balances by client. List disputed invoices separately and never recommend chasing them. Put anything over 45 days on the escalation list. Name every gap in the data. | Connectors. Today the billing coordinator uploads the Monday export. On the new platform, decide again: keep the upload, or add a read-only connection to the accounting system if the platform and IT support one. |
| Reference material. The firm’s credit terms, the escalation policy, the list of retainer clients and the summary template the partners read. | Permissions. Who can run the capability, who can edit it, which records it can read and where drafts are saved. Set in the new platform’s admin controls and in the accounting system, then checked with a real account. |
| Examples. Three past summaries the controller marked as the standard, with client names replaced by invented ones. | Execution. The coordinator starts each run by hand. A scheduled Monday run is a feature of the platform, with its own conditions, so the schedule is rebuilt and verified there. |
| Test cases. Twelve cases with expected results, among them a disputed invoice, a partial payment, a stale export and a client with a credit balance. Three are held back from development. | Packaging. How the instructions and reference files are bundled, installed and shared with the finance role, and how a new version reaches every user. |
| Ownership and review. The controller owns the capability and reviews every summary. The partners decide who gets contacted. AI prepares; people decide. | Testing. All twelve cases are run again on the new platform, and the results are recorded with the platform, model and version. The earlier pass does not transfer. |
The left-hand column is text and files in a folder the firm controls. The right-hand column is a set of decisions and configurations IT makes once per platform.
Why the right-hand column is real work.
Each item on the right is enforced by the platform, never by the instruction file.
- Connectors.Model Context Protocol gives AI applications a standard way to reach tools and information, and a standard protocol does not mean identical behavior in every workspace. Check the actual connection; supplied files are often enough. AI skills and MCP covers the choice.
- Permissions.A line that says “read only” in the instructions revokes nothing. Access controls live in the application, the server and the business system, so they are configured and tested again on every platform. AI skill security has the checks.
- Execution.The instructions describe the work; an application supplies the scheduler and the environment that runs it. Writing “every Monday” in a file starts nothing. How to schedule AI skills shows what to verify.
- Packaging.Under the open Agent Skills specification, a skill is a folder with a SKILL.md file that names the skill, describes when it applies and holds the instructions, beside its reference files. Installation and configuration still vary by platform, and plugin formats are specific to the host application. What is a SKILL.md file? and Plugins vs SKILL.md files explain both.
- Testing.A different platform can mean a different model, different tools and different defaults. The test set makes the move safe: the same cases, the same expected results, a fresh run. How to test an AI skill describes the method.
The adaptation checklist.
Copy it, fill in the brackets and run it once for each platform the capability goes to.
Platform adaptation checklist: [capability name]
- Confirm the method is complete. Instructions, reference material, examples, test cases and review rules sit in one maintained folder, with a version number. Anything that lives only in one person’s chat history is written down first.
- Strip platform wording from the instructions. Move every mention of a tool name, menu, file location or connector to a short platform note, rewritten for [new platform].
- Decide each input’s route. For every source the capability reads: upload, shared folder or live connection. Record what [new platform] supports and what IT approves.
- Set permissions where they are enforced. Who runs, who edits, what the capability can read, where output lands, and what it cannot do. Verify with a real user account, including one attempt that should be refused.
- Rebuild execution. State what starts a run: a person or a schedule. For a schedule, confirm the next run time, the missed-run policy and who can pause it.
- Package and install. Bundle the method in the form [new platform] loads, install it for the role and confirm how updates reach every user.
- Run the full test set. Every case, including the held-back ones. Record platform, model, version and result. Fix the harness first; change the method only when the owner agrees.
- Run it beside the current platform for [two] cycles. The reviewer compares both outputs on real work before the owner approves the switch.
- Retire or keep the old harness deliberately. Remove access and schedules on [current platform], or record that both stay in service and who maintains each.
Build for the move from day one.
Three habits make the right-hand column small. Keep the method in plain language, in a folder you control, under version numbers. Keep the test set beside it, growing with every miss; Build the test set that proves an AI workflow is ready shows how. And give the capability a named owner who outlasts the pilot, as Who owns an AI capability after the pilot? explains.
This is how Fully Powered builds: your platforms, your pace, your people. Every capability starts from the outcome and lands on the roles that do the work. We write the method to carry and name the adaptation for each platform up front, so your investment grows with every capability you add. That is how you 10X the AI you already pay for, on every platform you run. If you are still choosing a route for a piece of work, Your current AI platform, specialist software or a built capability compares the three.
Bring one workflow your team depends on and the list of AI platforms you run or are considering, and we will sketch its inventory with you.