Anthropic gave away the format behind roughly a million AI procedures this week. Seven notable models shipped in October alone. Here is how to tell which part of your AI stack is an asset and which part is rent.
Count the AI models that shipped in October.
Gemini 4 Argon on the 13th. DeepSeek V4.1-Flash on the 17th. MiniMax M3.1 Flash on the 19th. Qwen3.6 on the 21st. Kimi K3.1 Pro on the 23rd. GPT-6.1 Terra on the 26th. Gemini 4 Carbon on the 28th.
Seven, in one month, from six companies.
Now count the things you built with AI this year that would survive if your current tool disappeared tomorrow.
Here is the direct answer to the question in the title. You own exactly three things: your documented procedures, your written business context, and your own judgment. Everything else in your AI stack is rent. The subscriptions, the custom assistants living inside a vendor's product, the clever prompt you cannot find anymore, the automation built on a platform that just changed its pricing. All rent.
That is not a criticism of renting. I rent plenty. Rent is often the correct decision.
It is a criticism of calling rent an investment, which is what most of us have been doing for two years while telling ourselves we were building something.
And this week the industry made the distinction unusually clear.
Key Takeaways
- Anthropic donated the Agent Skills standard to the Agentic AI Foundation after roughly a million published skills and nearly 280 million installs across more than 70 platforms, which means the packaging format is now free and shared.
- When the format is open and the models change monthly, the durable value moves to the specific procedure you wrote down, not the tool you wrote it in.
- Research on small and medium enterprises finds that knowledge codification has a positive effect on performance, and that most SMEs do not give it enough attention.
- One long-term case study found codification effort can exceed the benefit, which is why you start with high-value, frequently needed procedures rather than documenting everything.
- The practical build is three documents: a business context brief, your two or three most repeated procedures, and the decision rules only you currently know.
The Problem
You have eleven AI subscriptions and no written procedures.
I am guessing at eleven. Go count. Include the one you forgot you were paying for.
Now here is the honest accounting. For each of those subscriptions, ask what remains if you cancel it. The custom assistant you spent a Saturday configuring? Gone. The prompt library inside the tool's prompt manager? Gone. The automation that strings four apps together? Gone, and the four apps still charge you.
That is the shape of the problem and it is not a moral failing. It is what happens when adoption is easy and documentation is not.
Setting up a new tool feels like progress because it produces an immediate result. Writing down how you actually do something produces nothing you can see on the day you do it. So the brain picks the tool, every time, and eighteen months later you have a stack and no asset.
I teach non-technical business owners and I run a community of around half a million entrepreneurs, and I would estimate the ratio is ten to one. Ten hours spent evaluating, configuring, and switching tools for every one hour spent writing down a procedure.
Which would be fine if the tools were stable.
They are not. Seven models in October. A vendor that cut its prices in half and got called a flop. A frontier lab that restricted its best model to vetted partners. A company that paused its own training runs and air-gapped them. If you build your operational knowledge inside any single one of those vendors, you have built on ground that moves every three weeks.
And there is a worse version of the problem, which is the one that actually costs money. The most valuable operational knowledge in your business is not in any tool at all. It is in your head. How you decide whether a prospect is a fit. What you check before a proposal goes out. Why you stopped offering the thing you used to offer. The exception you make for one specific kind of client and the reason.
None of that is written down, which means it is not an asset. It is a dependency, and the dependency is you.
The Evidence
Start with what happened this week, because it resolves an argument.
Anthropic donated Agent Skills to the Agentic AI Foundation, which sits under the Linux Foundation. Agent Skills is the format for packaging a repeatable procedure so an AI can execute it. The donation covers the standard along with its name, logo, GitHub organization, and the agentskills.io domain.
The numbers in Anthropic's announcement: roughly 1 million published skills and nearly 280 million installs, with more than 70 agent platforms having integrated the standard, including ChatGPT, Cursor, GitHub Copilot, and VS Code. Those are vendor figures, and at least one independent write-up noted that the announcement does not explain how they were counted, so hold the precision loosely. The specification stays public and free, and you do not need to join the foundation to use it.
Here is why that matters to a business owner who will never read a specification.
When a company gives away the format, it is telling you the format was never the valuable part. The value is in the procedures written in it. Anthropic co-founded that foundation with Block and OpenAI in December 2025, and engineers from Anthropic, Google, and OpenAI maintain the standard as individuals. Direct competitors are cooperating on the container because they are competing on everything else.
The container is now free. What you put in it is yours.
The same pattern showed up on Product Hunt this week from the other direction. Of the fifteen AI products on the board, the infrastructure layer was open source, Overmind, Paddock, and ContextGem among them, while the application layer stayed proprietary, TheySaid and NuPlay and Nuvio. Open plumbing underneath, paid vertical value on top. One auto-generated digest of the board described the shift as moving from generic model access toward specialized, context-aware systems.
Same conclusion, arrived at by the market rather than by a foundation.
Now the research on whether writing things down actually pays, because I am not going to tell you it does without checking.
A study of small and medium enterprises in India, based on a questionnaire to 428 personnel, found that knowledge acquisition and codification have a positive effect on performance. The same study observed that small and medium enterprises do not put enough attention on knowledge codification. That is a survey rather than a controlled trial, so read it as a correlation with a plausible mechanism.
And the counter-evidence, which is the more useful finding. A 2025 participatory case study published in Engineering Management Journal followed a medium-sized manufacturing plant and found that codification efforts were hampered by uncertain reliability of the information gathered. After fifteen years, many of the management team concluded that perhaps the effort to codify tacit knowledge was greater than any realized benefit. The authors suggested that on-the-job task completion under supervision may transfer tacit knowledge more effectively.
I take that seriously, and it changes the advice rather than negating it.
Documenting everything is a fifteen-year project with a questionable return. Documenting the three procedures that consume your week and currently live only in your head is an afternoon with an obvious return. The failure mode in that case study was scope, not the premise.
So: start narrow, start with what repeats, start with what only you can do.
What Actually Works
Three documents. That is the whole build.
Document one: your business context brief. One page, under 250 words, that you paste at the start of any AI conversation. What your business does. Who your customers are. Your offers and your actual prices. Your tone. The words you never use. The claims you never make. And one instruction that matters more than all the rest: tell the AI to ask rather than invent any fact you did not supply.
This single page eliminates the most common complaint I hear, which is that AI output sounds generic. It sounds generic because it was given nothing specific. You have been re-explaining your business from scratch in every new chat and every new tool, badly, under time pressure, and getting the output that request deserves.
Write it once. Keep it in a plain text file. Paste it forever.
Document two: your two or three most repeated procedures. Not all of them. The ones that eat your week.
A procedure is not a list of steps. That is the part people get wrong. A list of steps is something anyone could guess at. What makes a procedure yours is the judgment embedded in it: what you look at before you decide, the signals you watch for, the case where you did it differently and why.
The way to capture that is to talk through the task out loud while recording yourself, then hand the transcript to an AI and tell it to turn your description into a procedure with explicit decision rules, and to ask you about anything that seems to be missing rather than filling the gap. The gaps it finds are the valuable part. Those are the places where you have been operating on instinct you never articulated.
Document three: your decision rules. The questions your team always brings to you. For each one, how you actually decide. Written as if-then rules, with a clearly marked escalation line for the small set of decisions that genuinely stay with you.
This is the one that changes your calendar. Most of the interruptions in an owner's day are questions that have a rule behind them, and the rule has never been stated, so the question has to come to you.
Then three rules about how you store all of it.
Plain text or markdown, not inside an AI product. A file that will still open in ten years, on a machine you control, backed up where you back up everything else. If it lives inside a vendor, you did not build an asset, you decorated a rental.
Tool-agnostic language. When you review each document, hunt for anything that references a specific product's buttons, menus, or settings. Move those to a short appendix. The procedure itself should describe what has to happen, not where to click.
One named owner and three update triggers. Price changes, offer changes, and team changes. An asset nobody owns rots quietly until the day you need it, and then you discover it describes a business you stopped running fourteen months ago.
That is the system. Three documents, three storage rules, and a deliberate refusal to document everything.
Your Practical Steps
1. Count what you rent versus what you own, in writing. List every AI subscription with its monthly cost and annual total. Next to each, write what remains if you cancel. Do not give yourself credit for anything that dies with the tool. Most owners find the ratio uncomfortable, and the discomfort is the motivation for step two.
2. Write the business context brief first, because it pays off the same day. Here is the prompt:
[The Job]
Build the short business brief I paste at the start of any AI conversation.
This is for: me, saved somewhere I can reach in two seconds.
It matters because: I retype my business background several times a day and the AI still guesses wrong.
[The Background]
Here is what you need to know: my business is [WHAT YOU DO]. My customers are [DESCRIBE]. My main offers are [LIST WITH PRICES]. My tone is [DESCRIBE]. Words I never use are [LIST]. Claims I never make are [LIST].
Do not use: anything I have not supplied. Mark gaps for me to fill.
[The Deliverable]
Return: one brief under 250 words, ready to paste.
Must include: an instruction telling the AI to ask rather than assume any detail not supplied.
Optimize for: brevity and reuse.
[The Questions]
Ask me any questions you have.
3. Pick one procedure by return, not by ease. Rank your repeated tasks by hours per month consumed and by how much quality varies depending on who does them. Document the top one. Resist the urge to start with the easy one, because the easy one will not produce a result visible enough to keep you going.
4. Record yourself doing it and talk through the decisions. Not the steps. The decisions. What you look at, what would change your mind, the time you handled it differently. Then have AI turn the transcript into a procedure with explicit if-then rules and a list of questions about what is missing. Answer the questions. That round trip is where the real documentation happens.
5. Stress-test it against your messy cases, not your clean ones. A procedure that only handles the straightforward version produces confident wrong answers on everything else. Feed it the unusual cases you have actually encountered, add the rules that close the gaps, and write an explicit line for when to stop and ask a human.
6. Capture your decision rules by starting from the interruptions. For one week, write down every question your team brings you. At the end of the week, you have your list. Turn each into a rule, name the handful that stay with you, and hand the rest over.
7. Set the storage standard and the update triggers before you write anything else. Plain text, a folder you control, a backup, a named owner, and three triggers. Ten minutes of setup that determines whether this is an asset in two years or a folder of documents describing a business you no longer run.
Frequently Asked Questions
Is this the same thing as writing SOPs?
Close, with one difference that matters. A traditional SOP is written for a person who already shares your context. A procedure written for AI has to make the judgment explicit, because the AI has no context and will confidently fill any gap you leave. The result is usually better for your human team too.
Do I need to learn the Agent Skills format to do this?
No. Write your procedures in plain language in a plain text file. If you later want to load one into a tool that supports the standard, the conversion is mechanical. The thinking is the hard part and it is format-independent, which is exactly why the format was given away.
How long does one procedure take to document properly?
Budget ninety minutes for your first one, including the recording, the AI pass, answering its questions, and the stress test. Your second will take about half that. If a single procedure is taking more than two sessions, your scope is too wide and you should split it.
What if my business changes faster than I can document it?
Then document the things that are not changing. Your decision rules, your quality standard, and your customer context are far more stable than your tool stack or your tactics. Those three change on a timescale of years while your tools change monthly.
Should I document everything eventually?
No, and there is research arguing against it. The fifteen-year manufacturing case study found the effort to codify tacit knowledge exceeded the benefit when the scope was unbounded. Document what repeats, what only you can do, and what would be expensive to lose. Leave the rest to supervised practice.
The Close
Seven models shipped in October. The format for packaging AI procedures was given away for free to a Linux Foundation project maintained by engineers from three competing labs.
Those two facts tell the same story from opposite ends. The layer everyone is paying attention to is the layer that is converging, commoditizing, and changing hands every three weeks. The frontier is a rental market with excellent inventory and no loyalty.
Here is the thing. That is good news, if you know where to put your effort.
You do not have to keep up. You cannot keep up, and the people who do this for a living have mostly stopped trying. What you have to do is write down how your business actually works, once, in a plain file you control, in language that does not depend on any vendor's menu structure.
Then every model release is an upgrade instead of a migration.
Candidly, I believe this is the single clearest dividing line in small business AI adoption right now. Not who has the best tools. Who wrote anything down. The owners who spent this year collecting subscriptions have a monthly bill and a folder of expired configurations. The owners who spent it writing three documents have an asset that got more valuable every time a new model shipped, because the better the model, the better it executes what they already specified.
Same year. Same tools available to both. Completely different position.
Go write the first page.
P.S. If you want help turning your actual procedures into something an AI can run, with other owners doing the same work alongside you, that is most of what happens inside AI Insiders at $247 a month. But the context brief in step two is free, takes twenty minutes, and will improve every AI output you generate this week. Start there whether you ever talk to me or not.
About the author
Jonathan Mast is the founder of White Beard Strategies LLC, where he teaches non-technical entrepreneurs how to use AI to amplify the skill and experience they already have. He runs a Facebook community of roughly 500,000 entrepreneurs, the AI Insiders membership, and ongoing live training programs built on the Perfect Prompt Framework 2.0. He has built a seven-figure agency, lost it, and rebuilt it, which taught him the difference between the things you own and the things you are merely paying for.
