hq-mode: work through a whole task the lab's way¶

Type it at the start of a task, followed by what you want:

/hq-mode analyse the new run and tell me if the batch effect is gone

It never starts on its own. You turn it on, in Claude Code, Codex or Kimi Code.

What it does¶

First the agent reads one page of working rules. They cover when to ask you and when to find the answer itself, what it may do without asking, how it splits work between helpers, and how it writes its reply. A rule in your repo's own AGENTS.md wins over that page.

Then it picks the playbook that fits your task and copies its steps into its to-do list, so you can see what it will do and how far it has got. An agent with no to-do list puts the steps in its first reply instead. The playbooks:

When none fits, it designs one for the task, writes the steps down, and works through them.

It also follows a set of working principles. Each is a short page it reads when that principle applies, such as "check the real output, not a sign of it". A few were changed for science work. For example, before it trusts a cell count, it rules out data leaking between splits, batch effects, repeated samples and a filter that changed what the count is out of.

It also knows which skill to reach for along the way: the lab cluster skills before any ssh, unslop before you read its prose, and the planning skills such as /grill-with-docs and /to-spec for a new idea. A skill you start by typing its name stays yours to start. The agent suggests it rather than running it.

What it always asks first¶

Some steps wait for your yes, every time:

• submitting, running or cancelling a Slurm job, after it shows you the exact script in its reply, not a link to it;
• spending Edison credits;
• writing to or deleting from the lab's shared package store;
• force-pushing, or deleting data;
• merging a pull request nobody reviewed, unless you told it to land it;
• anything on a live service;
• making private work public;
• wet-lab steps;
• editing your own document beyond what you asked for.

This list wins over everything else you say in the session. "I'm going to bed, land it" lets the agent keep going on its own. It still stops at a Slurm job and waits for you, and you find the script ready in the morning. "Close the issue" is not "merge it": while the fix is not merged, the agent leaves the issue open and asks you.

Where the rules came from¶

They are the lab's way of working, drawn from many real sessions and written as "the user", so they apply to whoever types /hq-mode. Several playbooks build on pstack's poteto-mode, the principles come from pstack too, and the planning flow follows mattpocock/skills. Both are MIT licensed.