next level nerdingPi workflow
Pi workflow
Armin Ronacher’s Pi practice. David Ondrej hosts the interview. Armin Ronacher · GitHub · Pi
You do not type a slash skill to start. The agent already has read, write, edit, and bash. You type Unix. Add a skill file only when a tool earns it.
github.com/davidondrej/skills is the skills library — reference, not how you start. Do not install that repo to begin.
The video
David Ondrej interviews Armin Ronacher about Pi. Slides and brief command-line clips in roughly the first 13 minutes; the rest is talk.
Nine signals
The whole interview in nine boxes. Each opens the step in the map that explains it.
- 01
Four abilities, nothing else
Pi starts with read, write, edit and bash. No MCP and no sub-agents on day one, so you can see what the agent is doing.
How - 02
Search, don't paste
The agent finds what it needs with rg and the shell. You don't stuff the whole project into the chat.
How - 03
Add a skill only when a tool earns it
A skill is just a file the agent can read, like a CLI plus its README, or tmux. It is not a second framework.
How - 04
Explore on your own machine first
Armin works locally on a Mac for speed and control. He explores the problem before asking for a big batch of code.
How - 05
Heavy work goes to a box you control
When local is too hot or too small, he uses SSH to reach a Linux machine. Short-lived cloud sandboxes lack systems like Postgres and Kafka.
How - 06
One line of work, one worktree
Several jobs at once means one tmux session and one git worktree each. He also watches how hot the machine gets.
How - 07
You review before it ships
Output should be readable: Markdown, JSON or pipe output. Accepting code without reading it fails accountability.
How - 08
Write your private process down
Agents already know Git and Linux from training. Your own way of working isn't in there, so you feed it in.
How - 09
Augment people, don't replace them
Chat is still the interface. Software factories are a horizon, not a button. You stay the judge.
How
Words on this page
Fourteen terms, in plain words. Skip this if you already run coding agents from the terminal.
- Pi
- Armin Ronacher’s coding agent (pi.dev). It starts with only four abilities: read, write, edit, and bash.
- Harness
- The setup around the AI model that decides which tools it can use. Pi keeps its harness small so you can see what the agent is doing.
- bash
- The command line on Mac and Linux. Giving the agent bash means it can run the same commands you would type.
- rg
- Short for ripgrep, a fast command-line search tool. The agent uses it to find what it needs instead of you pasting the whole project into chat.
- MCP
- Model Context Protocol, a standard way to plug extra tools into an agent. Pi’s default is to skip it on day one.
- Sub-agents
- Extra agents the main agent hands work to. Also left out by default, so there is no second layer to keep track of.
- Skill file
- A file the agent can read that explains one tool, for example a CLI (command-line program) plus its README. You add one only when a tool earns its place.
- Codegen
- Code generation: asking the agent to write a large batch of code. Armin explores the problem first, then asks for it.
- SSH
- A way to log into another computer over the network and run commands there. Used here to move heavy work to a Linux box you control.
- Ephemeral sandbox
- A short-lived cloud machine that is thrown away after the job. The page’s point: it lacks long-running systems such as Postgres and Kafka.
- tmux
- A terminal tool that keeps several sessions running side by side. Used here to run several lines of work at once.
- Git worktree
- A second copy of the same project in its own folder, tied to the same Git history. One per line of work keeps parallel changes from colliding.
- Vibe coding
- Accepting what the agent writes without reading it. The page says it fails accountability, which is why human review is a hard gate.
- In-context
- Put into what the model reads during the task. Git and Linux are already in its training data; your private process is not, so you write it down and feed it in.
Follow it
What you type: rg, tmux, ssh, git worktree. The agent already has read, write, edit, and bash. Click a number to open that step in the map below.
Group 1 of 3
01–03setup once
Group 2 of 3
04–07every task
Group 3 of 3
08–10after the loop
The ten steps, with callouts
Ten boxes, left to right. Select a node. Arrow keys move between nodes.
Mostly a talking-head interview. Slides and brief CLI B-roll in the first ~13 minutes; the rest is spoken. No end-to-end issue → PR → merge run.
Keys: ← → between nodes. Home and End jump to the first and last nodes.
Try it, smallest version first
Five steps on your own machine. Install steps are Pi's own, on its site. The order and the starter prompt are FakeNerd's, not Armin's.
- 01
Get Pi
Install Pi from its own site. This page doesn't list install commands, so follow pi.dev rather than a copy here. Any coding agent you already run works for the next steps too.
pi.dev - 02
Start in one project on your own machine
Pick a small project folder you know and open the agent in it, on your own computer. No remote box, no MCP, no sub-agents, no skills repo. The four abilities are enough.
Step 01 in the map - 03
Ask it to look, not to build
Paste the starter prompt below with your own question. It tells the agent to search with rg and write what it found to a short Markdown note, without touching your code.
Step 02 in the map - 04
Read the note
Open notes.md and check it against the files it names. If it guessed, ask again. This is the review habit on the smallest possible output.
Step 07 in the map - 05
Then ask for one small change
Once the note is right, ask for the change and read the result before you keep it. Leave tmux, worktrees and SSH until one line of work is not enough.
Step 04 in the map
FakeNerd starter · look before you build
Search this project with rg and read only the files you find. Do not paste whole folders into the chat. My question: [what you want to understand, for example where the login form is handled] 1. List the files that matter and say why each one matters. 2. Explain how it works in plain words. 3. Write what you found to notes.md as short Markdown. Do not change any other file yet. I will read notes.md first, then tell you what to change.
FakeNerd's take
Copy the small harness first
Steps 01 and 02 cost nothing: four abilities and rg, with no MCP, no sub-agents and no skills repo to install. It works with whatever agent you already run.
Mostly said, not shown
Slides and brief command-line clips cover roughly the first 13 minutes. The rest is talk, with no issue to PR to merge run. The SSH box, the tmux and worktree setup and GitHub Actions are named, not operated on camera.
Review is the part to keep
Step 07 is the hard gate: read the Markdown, JSON or pipe output before anything counts as shipped. It's the cheapest habit here and the one most often skipped.
Built for terminal people
If tmux, ssh and git worktree are new to you, start with one line of work on your own machine. The remote Linux box and parallel lines can wait until one line is not enough.
The plain version under the others
On the Rauch page Ondrej calls Pi the closest small agent to fx, and Kun Chen's Firstmate runs on Pi among other harnesses. This page is the bare habits under those setups, with the fewest moving parts.
Related on this site: Rauch workflow, where Ondrej calls Pi the closest small agent to fx. Kun Chen's workflow, whose Firstmate runs on Pi among other harnesses. Company brain, which names Pi as a coding harness that can YOLO where a company-wide teammate can't.
Attribution
- David Ondrej's interview with Armin Ronacher (youtube.com/watch?v=SxuQs9GGYbk). The habits, tools, terms and reasoning are from this video, in our words. Each step in the map is marked as shown on screen, spoken, or both. The PostHog segment is a sponsor read, not part of Armin's stack.
- Pi is Armin Ronacher's coding agent; the link is from the video description. His blog and GitHub are linked so you can find his own writing.
- github.com/davidondrej/skills, David Ondrej's skills library. Linked for reference only; this page doesn't teach it and you don't need it to start.
- The ten-step path and its three groups, the map callouts, the glossary, the shown and spoken labels and the takes are FakeNerd's. Not sponsored by anyone on this page.