Posted September 24, 2026

Annoyed by Office 365’s forced shift to a subscription model and the hypocritical doomsday rhetoric of US AI tech giants I decided earlier this week to break on through — to the other side. Here’s what I replaced them with:
OnlyOffice (https://www.onlyoffice.com) — Open source, handles .docx, .pptx and .pdf natively. Free.
  Thunderbird (https://lnkd.in/ev_UzvVG) — Free, mature mail client.
  K-Drive (https://lnkd.in/eznsGrAh) — Swiss cloud storage, 3 TB, ~$5/month. Data stays in Switzerland.
  Mistral Pro (https://mistral.ai) with the OpenCase and OnlineKommentar connector for legal resources — European AI model, $15/month.
  Firefox (https://www.firefox.com) — Independent browser.
The migration took less than a day. So far the capabilities cover everything I need as a lawyer, judge, and board member. The other side is real — and it’s closer and better than I thought.

Mistral provided better legal content than Opus 5.5 faster (per September 2026). Both use the same MCP connectors to Legal Sources but Mistral found more relevant issues, wrote shorter and to the point. The only downer was the lack of native XML capability. Content had to be copied to Word where it had only basic formatting. If I wanted more elaborate formatting (eg for client facing documents) I had to run it through Claude for make it nicer without touching the content.

As I daily draft memos or do contract markup, I created some skills which create nice looking memos and do markups in my name in Word format and save them on Google Drive.

What actually breaks: the seam, not the drafting

A .docx file is a ZIP of XML files. When an AI generates that XML “by hand” instead of through a well-tested library, it makes small, predictable mistakes — and Word is merciless about them. In my case, three:

1. One element missing its namespace prefix — a single bare <vAlign>, invisible in any preview, well-formed enough that every XML check passed, and enough to make Word refuse the whole file.

2. Child elements in the wrong schema order — Word enforces a fixed sequence inside every properties block, and one pair out of order kills the document.

3. A duplicated element in one paragraph’s properties.

The weak point of Mistral is not drafting, it is the seams — the boring interfaces where open-source tools meet Microsoft’s file formats.

The fix: stop letting the AI improvise

The insight that solved it: separate the content from the container.

• The AI drafts content in Markdown — the master. That is what AI is good at.

• A small, deterministic Python script turns the master into the .docx — always the same way, always schema-valid, always in your house style.

• A built-in validator refuses to green-light the file unless it is actually clean. If it does not print VALIDATION OK, the job is not done.

The AI never touches the XML again. Neither do you. The script needs nothing but Python’s standard library — no Microsoft 365, no python-docx, no internet. It produces an A4 memo in a consistent corporate format (Arial, 18 pt title, To/From/Date/Status/Re box with shaded label column, confidentiality header, page footer, repeating table headers). Colours, fonts, margins and orientation are constants at the top of the file — five minutes to adapt to your own house style.

It also means your house style lives in your repo, not in a subscription.

The distribution problem — solved with a tool you already have

Another surprise was the lack of a download button. The solution is trivial: storage on Google Drive (although this is a bit ironical considerung the background of moving away from big US tech…)

-> chat for drafting, cloud drive for storage.

The template pack

I have put the whole loop online so you do not have to repeat my debugging afternoon:

1. memo_builder.py — the generator with the validator built in.

2. sample-master.md — a worked example so you can produce your first test memo in under a minute.

3. ai-memo-skill.md — a drop-in instruction file for your AI assistant. This file tells Mistral  the workflow: draft in Markdown, build with the script, validate, deliver to the cloud folder as a real .docx.

Getting started — in Mistral

The pack consists of one instruction and two data files. 

• The instruction (ai-memo-skill.md) tells the assistant how to behave when a document task appears: draft in Markdown, build with the script, validate, deliver as a real .docx.

• The data files (memo_builder.py, sample-master.md) are not instructions — the assistant must be able to read and execute them when needed. 

The fully automatic loop requires the paid plan of Mistral (15 Usd per month, 7 USD for students) as it needs Work mode (Vibe), because it can execute code and deliver files.

Work mode (Vibe)

1. Install the skill. Got to Context -> Skills → create new, with the content of ai-memo-skill.md. The description is the trigger — write it as “Load this when the user wants a memo or document as Word/.docx…”. The skill sleeps until its trigger words appear, so it does not clutter unrelated chats.

2. Store the data files where every chat can reach them: Context -> Knolwedge.  as Personal Knowledge files.

3. Test in a fresh, ordinary chat. Say: “Build the sample memo and deliver it as a .docx.” The assistant should find the builder, generate the document, run the validation, and only then deliver to the connected cloud folder. Open the file in Word or OnlyOffice. 

4. Adapt the house style. Tell the assistant what differs from the default — sender name and title, confidentiality marking, orientation, fonts, colours — and have it edit the MEMO defaults and house-style constants in memo_builder.py. Rebuild the sample and check. Iterate until “finalise this memo” reliably produces a file you would put your name on.

   Right-click → Save link as…
   https://adrianschaub.com/memoskill/memo_builder.py
   https://adrianschaub.com/memoskill/sample-master.md
   https://adrianschaub.com/memoskill/ai-memo-skill.md