← back

Running life admin as a data model

In September, a single price change on a house we were looking at cost me a full day. The figure appeared in 78 documents: budgets, comparisons, notes, and letters. I had to find and fix each one by hand, and I could not be sure I had found them all.

That day produced the rule behind everything in this post.

The rule

If a value can be computed from other values, it is not a fact.

Every project keeps its facts in one file, facts.toml. A fact is an atom: a price, a date, an income, or a deadline set by law. Everything computed from atoms (a monthly payment, a latest possible date, or a total) is a named derivation in Python. Neither Claude nor I type a figure from memory. Documents read their numbers from the generated output.

A price change now means editing one atom and regenerating.

Here is what that looks like in a small demo project. All the values are invented.

[fact.loan_rate]
value  = 3.6
unit   = "%"
as_of  = 2026-11-14
source = "Bank, first call"
status = "verbal"

[fact.hausgeld]
value  = 310
unit   = "EUR/month"
source = "Exposé"
status = "contradicted"
note   = "Exposé says 310; last year's statement says 345."

Every fact has a status

Each atom carries one of five statuses: confirmed, verbal, assumed, contingent or contradicted.

One command lists every assumed atom in a project. That list shows what a plan actually rests on.

$ uv run model/derive.py --status assumed
3 atom(s) with status assumed:

  equity                                      €95.000   own accounts, demo
  notary_rate                                  2,00 %   rule of thumb, notary + land registry
  repayment_rate                               2,00 %   our choice

Questions and deadlines

Two more registers sit beside the facts.

The demo project’s ranked questions, as the tool prints them. A question that blocks a date goes first, whatever it is worth:

  Worth To Question Asked
1 €150/month (≈€18,000 over 120 m) Bank Can you confirm 3.6 % fixed for 15 years in writing? 2026-11-14
2 €14,000 Agent Will the seller accept €398,000 in writing? never
3 €35/month (≈€4,200 over 120 m) Property manager Which Hausgeld is right: 310 or 345 a month? 2026-11-16
4 €3,000 Notary What will notary and land registry actually cost? never

The “never” column matters most. It separates the questions I decided not to ask from the ones I forgot.

A checker reads the live documents and confirms they still agree with the model. The model records retired values, such as an old price, so the checker can catch them if they reappear.

One machine, many projects

I built it for the house search first. A week later it became a shared toolkit, and every life-admin project now uses it: the household budget, a few government processes, and some family paperwork. A script creates a new project with the discipline already in place.

A schedule project reads across all the others and produces one agenda and one calendar.

Mail and todos

A Gmail poller routes new mail into each project’s inbox folder. A Claude session folds the inbox into the registers: a new date becomes an atom, a new question goes into the question register.

Todos started in Logseq. After a big harvest of todos in September, many turned out to rest on false assumptions, and Logseq could neither show me the list at a glance nor take a comment on one item. So every todo is now a GitHub issue in a private repository, on a project board with status, priority, and due date.

I correct things by commenting. An agent runs at 07:00 every morning, pulls new comments into the right inbox, fixes the fact, question, or deadline behind them, and replies on the issue. Family members use the same board: when a project needs a document from one of them, they get an issue and upload the document as a comment.

The archive

Scanned documents live in a self-hosted Paperless archive (I wrote about the scanning pipeline). Every project can search it, and a fact cites the document it came from by its archive id. “Passport (Paperless #N)” is a source someone can check. “Passport” is not.

What went wrong

The agenda once showed 34 overdue items. Several were already done. People closed the todo but not the deadline behind it. The rule now is that closing a todo closes its deadline.

The AI part

Claude wrote the toolkit, the pollers and the agent, and it does the daily bookkeeping. The instructions forbid typing a number from memory, so every figure it writes comes from the model.