Running
[ work ] llm-wiki · running
A repository of 69 rules, each written from an incident and carrying that incident with it, read by every agent before it does any work. Built and run by one person and a team of agents — which half did what is named further down.
Before
A lesson learned in one session lived in that session's memory, and a subagent never receives that. So the same mistake stayed available to every agent that had not personally made it.
Now
The rules live in their own repository and arrive before the first token of work. The planner and the critic read them before proposing or reviewing; the writer reads the one page that governs what it writes.
$ agent start --role planner
rules.md — how work is done
------------------------------------------------
R24 Never state a record as fact without
checking it.
R30 Do it yourself before handing it back.
R47 Align once before starting, then work
without asking.
69 rules · 11 pages · injected before the
first token of workwhat the thing does
A mistake is made once, and then it belongs to every agent.
One direction only — a rule enters where a human approved it
The planner and the critic read the rules before they propose or review — not as a checklist afterwards, when the work is already shaped.
Drafting is automated. Approval never is. A system allowed to write its own rules will write rules it invented, so that one step stays human.
A scheduled lint reads every page, counts whether each claim still carries its source, and reports the axes it cannot check itself.
69
rules, never renumbered
11
pages
87
citations, lint green
79
commits, each one read
[ 01 ]
Of twelve rules it proposed from raw transcripts, two already existed — rediscovered independently, from the incidents rather than from the page. That is the clearest evidence the detector reads behaviour instead of echoing us.
[ 02 ]
Everything ruled out keeps its own auditable page, with the measurement that ruled it out. When the same proposal comes back — and it does — it meets its own evidence instead of a vague memory.
[ 03 ]
87 citations, zero open issues. The lint also names the two things it cannot check, so a green run is never mistaken for a clean bill of health.
Both halves are named on purpose. The interesting part of this one is the seam — what is worth automating, and the one step that is not.
The human
The agents
It is a list of rules that the AI assistants here read before they start any work.
Here is the problem it fixes. When I catch one of them making a mistake and correct it, that correction only lives in that one conversation. The next assistant I ask for help is a fresh one — it never heard the correction, so it is free to make exactly the same mistake. And it does. Every one of them ends up learning the same lesson the hard way, over and over.
So the corrections got written down in one place, and now every assistant reads it before it does anything. There are 69 rules so far. One of them says: never state a record as fact without checking it — written the day one of them told me something confidently that turned out not to be in the file it claimed to have read.
Each rule carries the mistake it came from. That part matters more than it sounds: a rule with a real story attached is hard to wave away later, and easy to check against.
What it is not: a manual, or a description of how anything works. It only contains things that went wrong once.
How a rule gets drafted, what the weekly lint cannot check, why approval is the one step that is never automated — ask, and the answer goes up on this page.