subaud

the guide

Working this way

How we start a project and build on it, from a set-up machine to a release, in eight stages. Open one to see what happens, what to do, and which pieces of the kit it uses.

the stages

From a set-up machine to a release

1

Keep a machine up to date

kempt brings each new release

kempt installs the tools, their skills and your settings from one file, and shows every change before it makes it. Connect sets a machine up with one line; after that, kempt keeps it matching the file and brings each new release to it. It's one of the few tools you run yourself. Keys go in through creel from the start.

Terminal
kempt update

Uses: kempt · creel · pi-provider-guard · pi-auto-reload

2

Start a project

one proj session per piece of work

Every piece of work is a proj session, named ‹project›/‹work›, in the project's folder with the agent in it. A workspace is a folder that holds related repositories, so one session can change them together; each branch gets its own worktree inside it. Each session has its own scratch pad for things you'll need in ten minutes. We usually start the session ourselves, from proj's page in the browser: pick the project, name the work, choose the agent. Once you're in the session, tell the agent what the work is.

To have the agent open one for you instead:

Say to your agent
Open a proj session for ‹project› so I can work on ‹what› there.

Uses: proj · scratch · a git worktree per branch · pi-tmux-bridge

3

Plan it

decisions on diazo's page

Any change you'd like to see before it's built starts on diazo's page: the agent says back what it understood, posts the decisions as cards, then a spec and plans you approve. A one-line fix with one obvious answer skips this; just ask.

Say to your agent
Plan this with diazo: ‹your idea in a sentence›.

Uses: diazo · brainstorming · writing-plans · one question at a time

4

Build it

a worker builds, a reviewer checks

A worker subagent builds the approved plan in its own worktree, test first. When it's done, a reviewer from the other model provider reviews the whole change, and the agent checks each finding before acting on it.

Say to your agent
Build the approved plan.

Uses: executing-plans · pi-subagents · test-driven-development · a reviewer from the other provider · receiving-code-review · systematic-debugging

5

Review the writing

galley

READMEs, pages and specs are reviewed in galley before they ship: you mark up the draft, the agent revises, and you approve when it reads right.

Say to your agent
Let's use galley to edit ‹file›.

Uses: galley · reviewing-a-document

6

Ship and land it

a pull request with auto-merge

A finished branch becomes a pull request with auto-merge on, and merges once its checks pass. Long builds and deploys run in a tmux window you can watch, and the agent waits with a wakeup rather than inside a command. Once a release is out, kempt brings it to your machine; an npm package goes out through imprint.

Say to your agent
Open the pull request with auto-merge on.

Uses: a pull request with auto-merge · running-long-jobs · pi-wakeup · imprint · kempt

7

Run many sessions

muster and channels

Several agents work at once, one per session. muster lets them tell each other when one's work lands, or ask another to review; channels wakes the session that a message is for. When a Claude subscription runs out, the session switches accounts and carries on.

Say to your agent
Tell the ‹session› session: ‹message›.

Uses: muster · channels · proj · pi-schedule · pi-usage · switching Claude accounts

8

Keep it tidy

casebook, sift, cull, muda

casebook gives every branch, pull request and worktree a decision; sift trims the instruction files your agents read; cull cuts tests that don't earn their place; muda finds waste in the delivery pipeline.

Say to your agent
Use casebook to go through my stale branches and recommend what to do with each.

Uses: casebook · pi-casebook · cleaning-up-branches · sift · cull · muda · remembering