Rulebook
Getting an expert's unwritten rules out of their head
- Clients whose rules live in their heads. Client work, so nobody is named.
- 2 September 2026
- What the build currently believes, one claim per line
- The expert's corrections, collated per person
- None. The link is the key.
The hardest part of building for a trade business usually isn’t code. It’s the rules the owner knows and has never written down.
Rulebook doesn’t ask the expert what they know. It tells them what the system currently believes about their business, one short claim per line, and lets them mark each one right or wrong. A wrong claim is the useful one: it draws the correction, with the real rule and its exceptions attached.
- Rounds on a phone. Each round is a single page the expert works through top to bottom, with a note field on every item.
- Claims, questions and choices look different, and the page refuses to build if an item’s control doesn’t match what it is. A right/wrong tick on a question can’t confirm anything.
- Answers collate as they come in, and the corrections fold straight into the build’s rule book.
First real use, with a cabinetmaker: 39 items in 14 minutes, with five corrections and seven additions. The next round, on the version described here, got 30 of 34 answered in twenty minutes.
Stack
- Claude Code
- Writes each round with me from what the build already believes, then folds the corrections back in.
- HTML
- The round itself. One page, no framework, and it loads nothing from anywhere else.
- FastAPI
- Serves the round and saves every tap as it happens.
- Python
- Signs each link, checks a page before it builds, and collates the answers.
- Cloudflare Tunnel
- Gets the page from my home server to the expert's phone.
- Resend
- Emails the link, and tells me when someone starts, taps Done or the round closes.
Build log
-
The answer collector had been fetching its keys on every run, and helped run my password manager's daily allowance dry. It now reads them once and stays running. Read →
-
First client round on the phone. The cabinetmaker answered 30 of 34 in twenty minutes, and his corrections went into the rules that night.
-
I ran rounds on my own projects with false claims planted about how my systems work, and ticked every one Right. On my own projects, a claim about the machine is now checked against the code, not put to me.
-
A program can write a round now, not only me. Sketch to CAD sends its own questions as a link, and tapping Done redraws the job. Proven on my own taps.
-
A round on myself counted notes like "not following" as answers. Items now carry a Not sure button, and an unsure is a fault in the item, not the person.
-
On the open internet with no login. Each link is signed for one round and one person. Two people can answer the same round, and where they disagree both answers are kept.
-
Started. A cabinetmaker marked up 39 items in 14 minutes, with five corrections and seven additions. Some items were questions I'd given Right and Wrong boxes, and those ticks confirmed nothing. The page now refuses to build that way.