Instructions our agents read when a task matches: how we plan, build, test and review. The working-method skills below aren't published yet. Several are adapted from obra/superpowers (MIT). The tools' own skills (diazo, galley, cull, sift, muda, imprint) come with each tool.
Paste in a terminalSay to Claude Code or pi
From idea to merged codeThree skills, one after the other: brainstorming → writing-plans → executing-plans
what
The path from an idea to merged code. brainstorming turns an idea into an approved design, one question at a time, on diazo's page when the design is big. writing-plans turns the approved spec into a plan with exact tasks, files and checks. executing-plans has a worker subagent build it, test first, then a reviewer from the other model provider reviews the whole change.
why
Each step needs your approval before the next, and the change is kept to the correct size for the ask: as small as correctness allows, with what it leaves out stated.
Write the test first, watch it fail, write the smallest code that passes, watch it pass. One test per behavior the plan asks for, or one that reproduces the bug actually seen.
why
If you didn't watch the test fail, you don't know whether it tests the right thing.
Find the root cause before fixing a bug, in four phases: read the real logs and state, compare with code that works, test one idea at a time, then fix at the cause with a test that reproduces it.
why
A fix aimed at the symptom usually moves the failure somewhere else, and each guess that half-works makes the next one harder to read.
ask
Say to your agent
The build is failing. Find the root cause before changing anything.
How the agent handles a review's findings: check each against the code first, fix the ones that show the approved behavior failing, push back on wrong ones with evidence, and note the rest as follow-ups.
why
A review is evidence to check, not a list of orders. It is finished when the approved behavior works and nothing regressed, not when a reviewer finds nothing.
ask
Say to your agent
Go through the reviewer's findings and tell me which ones you'll fix.
A command that only needs to finish (a deploy, a build, a test suite) runs in its own named tmux window. You can watch it live; the agent checks its log when it's done and closes the window.
why
You see the job in the status bar, and the agent spends no turns watching it.
ask
Say to your agent
Run the test suite in a tmux window I can watch.
(not published yet)
rememberingSkills
what
There is no memory store. When you say "remember this", the agent finds the file the rule belongs in (your global instructions, a repository's CLAUDE.md, an issue) and proposes the edit for you to approve.
why
Things worth keeping live in a file that is read in full, versioned and reviewable, where you can see and change them.
ask
Say to your agent
Remember this: ‹rule›.
(not published yet)
cleaning-up-branchesSkills
what
How to decide which git branches and worktrees are safe to delete, including branches whose pull request was squash-merged.
why
A squash-merged branch looks unmerged to git's usual checks, so finished work can look unfinished.
ask
Say to your agent
Clean up the merged branches in this repository.
(not published yet)
pi-subagent-dispatchSkills
what
How a pi session on our machines picks the model and provider for each subagent, and what to do when a dispatch fails.
why
It carries out the model rules (who builds, who reviews whom) without the session having to work them out each time.
ask
(works by itself once installed)
(not published yet)
driving-chromeSkills
what
How a session clicks, fills and reads pages in your own Chrome, with your logins, through the Playwright MCP server.
why
Signed-in sites need your real browser; you are never asked to log in to a separate one.
ask
Say to your agent
Open the sign-up page in my Chrome and fill in the form.
(not published yet)
using-the-toolsSkills
what
Picks the subaud tool a request calls for, or none, and hands off between them. It is installed with every subaud tool.
why
A one-line fix needs no tool; a change with a choice in it needs diazo. The agent reaches for a tool only when the request calls for it.
A worktree is a second checkout of the same repository in its own folder, on its own branch. Every branch an agent works on gets one, under the workspace's .worktrees folder; the main clone stays on its base branch.
why
Several sessions can work on one repository at once without switching branches under each other.
The agent asks one clear question, with the options, their pros and cons and its recommendation, then waits. It decides the ordinary calls itself and lists them so you can reverse any.
why
Each question gets a real answer, and you only decide what is genuinely yours to decide.
ask
Say to your agent
Ask me one question at a time, with your recommendation.
Keys through creel, never in chatHabits and tricks
what
Any key, token or password goes through creel's masked box. Before asking for one, the agent says what it is, what access it grants and what it will be used for.
why
A secret typed into the chat is in the conversation; one typed into creel never reaches the agent.
ask
Say to your agent
I need to give you ‹KEY_NAME›. Ask me for it with creel.
When a branch is finished, the agent pushes it, opens a pull request and turns on auto-merge, so it merges by itself once its checks pass.
why
The agent doesn't stop to ask you for each merge, and the repository's checks still decide whether it lands.
ask
Say to your agent
Open the pull request with auto-merge on.
(a habit; no package)
Switching Claude accountsHabits and tricks
what
When one Claude subscription runs out, the session switches to another with /claude-account, the accounts panel in our pi-claude-bridge fork, and its subagents follow.