Standing instructions

Standing instructions are the rules agents follow without being told each time: where the data lives, what never changes, how results are reported. ArkenOS reads them from a few places, each with its own reach. Put each rule in the narrowest place that covers it.

Why it matters: without them, you repeat the same rules in every chat, and an agent that wasn't told guesses. A guess like training on the raw ticket export, duplicates and all, costs a whole run to discover.

Where instructions go#

The workspace AGENTS.md#

A file called AGENTS.md at the top of the workspace folder goes into every agent's instructions: Optimus, the specialists it brings in, tasks it hands off, and agents replying in channels. Use it for the rules of the project itself.

# Support-bot fine-tune

We fine-tune an 8B base model with LoRA to answer support tickets.

- Training data: data/clean/. The raw exports in data/raw/ have duplicates; never train on them.
- Held-out set: evals/heldout-tickets.jsonl. Never train on it, never edit it.
- Every result: three seeds, mean and worst seed, beside the base model (71.4% resolved).
- Ask before any cloud run over $50.
- Results go on a page under Experiments/, numbers first.

• Create it by asking. “Write an AGENTS.md at the top of the workspace with these rules: …” Optimus writes it for you. You can also create it with any editor in the workspace folder. It stays out of the Pages list.
• Changes reach new chats. A chat reads the file when it starts, so start a new chat (or a new task) after editing it. If there's no AGENTS.md, a CLAUDE.md in the same place is read instead.
• Keep it short. Every agent reads it on every session. A page of rules gets followed; ten pages get skimmed. Move long procedures into skills.

The workspace home page#

Each workspace starts with a home page, its README.md, holding the workspace name and description. When agents reply in a channel, an issue or a page comment, they read the first 4,000 characters of it as the workspace brief.

Make the top of the home page say what the project is and what it's for, in a few lines. That's what an agent who's tagged into a channel sees first. To give agents a separate brief, put it in .arkenos/README.md; when that file exists, agents read it instead.

One agent's instructions#

Each agent has its own AGENTS.md, shown as System prompt on its profile card. It's read at the start of every task that agent does, and the agent adds to it as it learns. Use it for rules that belong to that agent's job:

See Edit an agent to open and change it. For the agent's tone rather than its rules, change its personality.

Tell it in chat#

The quickest way to set a rule is to say it: “Remember that we always report the worst seed.” The agent saves it as a memory note, so it carries over to later chats. Use this for preferences that come up in the moment. Write the workspace AGENTS.md for rules every agent must follow from day one.

Pick the right place#

• A rule for the whole project (paths, budgets, what's off limits): the workspace AGENTS.md.
• What the project is about, for agents joining a conversation: the top of the home page.
• A rule for one role (how the checker checks): that agent's system prompt.
• A procedure for one job (how to run the eval): a skill.
• A one-off limit for this task: the message itself. “Keep the learning rate as is” belongs in the ask.

When guidance is ignored, check where it lives before rewording it. A rule in one agent's system prompt reaches that agent alone, and a procedure buried in a long AGENTS.md gets lost. Moving it to the right place usually fixes it.