FAQ
The longer answers
What your AI provider sees when a skill runs, who decides what is confidential, and what to ask the vendor your company already signed with. The short version is on how it works.
01
Your data and your AI
The only system that reads your documents is the AI you point at them. MOS is files on disk — no account, no server, no copy of anything.
Does MOS get a copy of my files?
No, because there is nowhere to send them. Most tools ask you to upload your documents to them. MOS is a folder of instructions and templates you download, and it has no account, no server and no database.
Then what reads my documents?
The AI you run those instructions with. If that is a cloud AI — Claude, or whichever one your company uses — then the contents of every file it opens are sent to that provider, the same way they would be if you pasted the text into a chat window yourself. That is how cloud AI works, with or without MOS.
So the question worth asking is not whether MOS is safe to use with confidential material. It is which AI is reading your documents, and what your company's agreement with that provider says.
What is the difference between uploaded and sent as a prompt?
Uploaded
Persisted, ingested, indexed, sitting in a vendor's database — what a SaaS medtech tool does with the documents you hand it. MOS does none of this. There is nothing to upload to, and no index anywhere.
Sent as a prompt
Transient inference by the AI provider your company already chose and already contracted with. This does happen. It happens under an agreement you already signed, and MOS is not a party to it.
What does that mean in practice?
What your AI provider sees
Whatever files a skill opens while it runs, as prompt text. Retention, training and data residency are governed by your existing agreement with that provider — not by MOS.
Who decides what is confidential
You do. Every company has its own classification scheme and AI-use policy, and MOS is not equipped to override them. MOS builds no redaction proxy and no content filter: a filter that is right 95% of the time is worse than none, because people trust it.
If nothing may leave at all
Skills are plain Markdown instructions, not a hosted service, so nothing in the design requires a cloud model. Against a local model, nothing leaves the machine except the public-database searches described below, each shown to you first. Running skills on a local model costs you output quality, and MOS says so rather than presenting it as free.
How much of my context does a skill read?
What the job needs, from the folders your map points at. Skills read whole files by name rather than pulling fragments out of a search index, so what the AI saw is a list you can read afterwards. A skill that needs something outside the map asks for it by name and records your answer.
These are rules the AI follows, not a lock on your files. They make what gets read predictable and visible; they do not make a mistake impossible. A company that cannot tolerate a single one should set file permissions on the machine — an operating-system job, not a job for a convention.
Which public databases do skills search?
Five skills search government records, because a predicate or a coverage policy found from memory is worse than none: it looks researched. The requests go from your machine straight to the agency, not through MOS, and carry only generic terms. Never your product's name, its indications, its claims, its description or any file's contents. Every request is shown to you before it runs, and logged in the document it produces.
FDA
openFDA, FDA's device databases and fda.gov. Used by product code, precedent search, differentiation and indications strategy. Sent: generic device words, product codes, regulation and submission numbers, competitor names.
Office of the Federal Register
The eCFR, for regulation text. Used by product code and precedent search. Sent: regulation numbers.
National Library of Medicine
ClinicalTrials.gov, used by indications strategy: conditions, generic device words, competitor names. Clinical Tables, for billing and diagnosis codes, used by reimbursement: generic procedure, device and condition words.
CMS
The Medicare Coverage Database, used by reimbursement. Sent: no search terms. The skill downloads Medicare's complete policy lists, searches them on your machine, and reads national policies by number.
Differentiation also offers a web search for what competitors publish. It runs through your AI tool's own web search, under that provider's terms, and only if you say yes.
Questions for your AI vendor
Before pointing any AI at company files, check the terms your company already agreed to: enterprise tier rather than consumer, training on your data disabled, retention at zero or a bounded window, a BAA if patient data is ever in scope, and data residency where EU MDR makes it matter. These are questions for that vendor. MOS cannot make guarantees about a contract it is not part of.
02
Using MOS
What you need before you start, and where the outputs sit once you have them.
Do I need to know Git?
No. MOS downloads as a ZIP, from the link in the community: download, unzip, drop it into your folder. Git is there for people who want it.
Do I need Python?
Yes. The shared scripts that write Word documents, search public databases and size studies are Python. Setup checks for it and asks before installing anything.
Do I need an internet connection?
Your AI needs one if it is a cloud AI. Beyond that, five skills search public databases, and without a connection they say they cannot reach the service.
Which AI do I need?
One that can read and write files on your machine. A browser chat window will not work — it cannot open your context, so a skill run there is a template with no company in it. What you need is a coding agent or AI code editor, despite the name: Claude Code is what skills are tested against. Skills are designed and tested on Claude first; where one fails elsewhere, that is a bug in the skill. MOS does not ship an AI and does not sell one.
Does this replace my quality system?
No. Skill outputs are drafts in outputs/ — work in progress, not records. They become part of your quality system when a qualified person reviews, approves and imports them, on your own terms. The AI is a tool; the author is the person who approves the work.
What does it cost?
Nothing for MOS — content is CC BY 4.0 and code is MIT. You pay your AI provider. An AI that works in your files is not always the same product or plan as a chat window, so check what your company's plan covers before you count it as free.
Do I have to reorganize our folders?
No. You map your existing folders to the MOS vocabulary once, and the skills work against your layout from then on. Starting fresh, setup scaffolds a layout for you.
A question that is not answered here is worth asking in the community — that is where MOS gets argued about and improved.