Rules
Write your house rules once, and every chat in the project starts out knowing them. A task's workers do not read them: from a rule, a task gets only the model choice.
A rule never fires. It is simply true in every chat it reaches.
- security fixes always run on the strongest model
- ask before starting anything big here
- doc comments in full sentences
A rule is a standing order with nothing to wake it. It has no time, no rhythm and nothing to look for, so it is never checked, never fires and costs nothing. Its whole job is to be in front of the work: every conversation in its reach reads it on every turn, under the heading Standing orders, told that these are your conditions and not suggestions. On every list it says holds where other orders say when they wake.
Try it in CodeAF
- Where
In a conversation in your project.
- Type
/standing security fixes in this repo always run on openai/gpt-oss-120bThen press Enter.
- You see
A card reads
wants to keep a rule, withwhere · for this project. Press 1 to keep it.
What a rule can hold
Anything you would tell a new colleague on their first day. Conventions: "we use tabs", "prefer small commits". Boundaries: "ask before starting anything big", "never touch the public API". How it writes to you: "keep replies short".
A rule can also decide who does which work. With the rule above standing, ask for a security fix in plain words and the chat hands the task to openai/gpt-oss-120b, while other work keeps the usual crew. The rule acts through the chat that hands the work out, so a /task you type yourself still gets the usual crew.
Reach, and the one place it does not apply
Said in a chat, a rule covers the whole project. "Everywhere" or "in all my projects" lifts it to every project on the machine, and "just this chat" keeps it to one conversation. The card names the reach before anything stands.
If you say a rule and no card comes, CodeAF may have kept it as a note. Send it as /standing <words> and it is always read as a rule.
Things to build
- A security model policy: "Security fixes in this repo always run on openai/gpt-oss-120b." Ask for a security fix in plain words and that task runs on it, while other work keeps the usual crew.
- A cheap lane for chores: "Docs-only changes in this repo run on z-ai/glm-5.3-flash": the chat hands those tasks to the cheap model when it starts them for you.
- Replies that fit your screen: "In all my projects, keep replies to five lines unless I ask for more." The card reads
where · everywhere, and every conversation on the machine reads it on each turn.
Go further
Standing orders: Say once what should stay true, and CodeAF keeps it without being asked again.
Coming soon: describe the org you want and CodeAF builds it: teams, managers, budgets, standing orders.