Skip to main content
Every team has rules that live in one person’s head. “Always check the queue depth before launching.” “We never merge without re-running the parity suite.” Workflow memory is where those get written down once, in a form a stranger’s agent can act on.
Workflow memory is behind a per-user feature flag. On accounts outside the rollout the probe_procedures MCP tool is not advertised and the /set-rule skill is not shipped — that is the gate working, not a broken install.

​
Writing one

Two steps, always. The first structures what you meant; the second stores it.
preview writes nothing. It classifies the prose into a situation and shows you the draft that would be stored. Read it, edit it, then declare:
--related names the clause the relation points at.
Nothing is written until a person confirms it. The point is to capture what someone meant in a form a stranger’s agent can act on — not to transcribe their words. Show them the structured reading and ask.

​
Publishing

A declared rule is yours until it becomes the team’s. A rule becomes visible team-wide when a second person declares the same thing — or immediately, on your own authority:
Publishing is unilateral and attributed to you. Withdrawing exists because of that: a rule the team turns out to disagree with should come down without deleting the clause and the evidence behind it.

​
Reading them

With --query, the store classifies first and may legitimately return nothing: an unclassifiable query serves zero rules rather than guessing a situation and handing you rules for the wrong one. The situations the store is organised around are the moments where being wrong is expensive: launching a run, opening a PR, deploying, processing a dataset, running an eval, reviewing code, provisioning infrastructure, reproducing an experiment, debugging a failing run, recovering from an incident, claiming work is done.
A rule that the classifier cannot place lands in a misc bucket and surfaces explicitly. It is never dropped.

​
Zero rules is four different answers

The response distinguishes them in completeness, because “no engine wired”, “this workspace never opted in”, “nobody seeded the vocabulary” and “the team has not written one down” are otherwise identical bytes — and only the last one is an answer.

​
For agents

The rule for agents is check before the step, never after: before an irreversible or shared-consequence action, and on arriving in an unfamiliar workspace.

Agent skills

How /read-rules and /set-rule fit into a session.