A talk

Building fast is solved. Deciding well is not.

A talk on a career break, 2026.

Robert Blust · Software Engineer & Architect

rests onexperience Building fast is solved. Deciding well is not.

The claim

The constraint moved.

Decide well over build fast: “AI moved the constraint from building quickly to deciding correctly; the decision is the work.”

A value in the model, quoted as it stands there.

rests onvalue Decide well over build fast

The break

One question, four months.

June to September 2026: what to do next, tested by building in the open.

It ended with a decision, not only an offer.

rests ondecision Four months to find out what I want to do next decision An 80% architect role in a product company, chosen over pay

The first half

Building fast, measured.

23public repositories in 3 organizations

5,065commits on default branches

1,640merged pull requests

272releases

9live hosts

91%of non-merge commits co-authored with Claude

One person. As of Sep 24, 2026.

rests onstrategy Model-First Strategy skill Agentic AI development

The curve

The curve.

Commits per day; each tick marks the day a repository was born.

rests ondecision The two products are built in the open

The cost

Building is not the expensive part any more.

about CHF 650for the subscription in four months, paid

about CHF 9,000of compute at API list price

about 16 billiontokens, most of them context read again

It rests on a few providers’ prices and terms. The knowledge is portable; the speed is not yet.

What the same result would have cost a Swiss team: blust.ch/talks/deciding-well/cost

rests onstrategic objective Every service I pay for is sized to what I use kpi Plan Fit

The second half

Where the time went instead.

95specs

105plans

5phases, each behind a gate

Shape · Spec · Plan · Implement · Integrate — the Owner merges.

rests onprocess Delivery rule The Owner's word merges

The decisions

Taken, revised, dropped.

1,360decisions written down in 95 specs

72revised later

30dropped or parked

1,258still standing as written

1revert in 5,065 commits

12repositories in one release wave, replaced the same day

An agent had to read them out of prose. Next, decisions become part of the model.

rests onprocess Deciding

The answer

The answer: one graph.

CompanyGraph: vision, strategy, roles, processes, rules and their evidence as one graph, for people and agents.

Plain Markdown in git, served over MCP — no provider owns it.

rests onconcept Mental Model experience CompanyGraph

It runs

It runs.

blust.chthe mental model of Robert Blust, a company of one
site · MCP server · chat

companygraph.iothe mental model of CompanyGraph, describing itself
site · MCP server · chat

guestgraph.iothe mental model of GuestGraph
site · MCP server · chat
and the product in code: the open-core engine and the Apaleo connector

Three mental models, each in git, each served three ways. Ask a chat a question its model can answer.

CompanyGraph and GuestGraph are ready for scrutiny: open core, validated in the open.

rests onsurface blust.ch website surface mcp.blust.ch MCP server surface chat.blust.ch chat

What comes next

Deciding well is the work.

Building fast is solved. Deciding well is not.

My proposal: decide well with a mental model and agentic AI. Validate it with me.

blust.ch · companygraph.io · guestgraph.io — from October, an IT architect role, chosen over better-paid engineering and lead roles; from 2027, education in how AI is led and governed.

rests onstrategic objective Whoever decides about me decided from the model

Where to start reading.

The argumentThe valuesblust.ch/principles

The proofEvery claim, with its evidenceblust.ch/model

The answerEverything a company knows, one graphcompanygraph.io

Who decides whatblust.ch/processes

The career as a ledgerblust.ch/timeline

What a team would have costblust.ch/talks/deciding-well/cost

Where the model is publishedcompanygraph.io/surfaces

Two ideas, put up for scrutinyblust.ch/ideas

rests onvision One model, true everywhere

Building fast is solved. Deciding well is not.

The constraint moved.

One question, four months.

Building fast, measured.

The curve.

Building is not the expensive part any more.

Where the time went instead.

Taken, revised, dropped.

The answer: one graph.

It runs.

Deciding well is the work.

Where to start reading.

Sprecher-Notiz