pstack, peeled
pstack is a free set of written instructions that makes AI coding agents work carefully: understand first, try ideas, prove the work, and only then say "done." This guide peels it open one layer at a time, in plain words, with things you can tap and play with.
pstack was made by Lauren Tan, who goes by poteto online. Hence the potato. This is a fan-made explainer, not made by or affiliated with Lauren, Cursor or Anthropic. Facts come from pstack's own files and Lauren's public posts (sources at the bottom).
- An AI model only knows what's in its for this conversation. Nothing else.
- It can't touch anything. It can only write text, including like "open this file." An app such as Claude Code or Cursor carries those requests out.
- pstack is a folder of written instructions called . The app copies the right one into the AI's memory at the right moment.
- Those instructions make the AI act like a careful senior engineer: check, try, prove, and only then report.
Six players, one kitchen
Everything on this page involves the same six players. We'll use one comparison the whole way through: a kitchen. Tap any card for the full story.
The one thing to remember: the AI is a brilliant cook who can't leave the stove and can't remember anything except what's written on one notepad. pstack decides what gets written on that notepad, and when.
Watch one request travel
Press play. The little potato carries your request through the system, then carries the AI's requests and the answers that come back. The yellow strip under the chart always says, in plain words, what's happening right now.
Before you press play
Sizes are rough estimates, in , out of a typical 200,000-token memory.
The AI writes when its box is lit up. Its requests (called tool calls) show up here.
A skill is a recipe card
A skill is a short written instruction saved as a file. When it's needed, the app copies the card's text into the AI's memory and the AI follows it. Nothing gets installed in the AI. Take the card away and the AI forgets it.
"Your words carry the intent, and the skill carries the rigor."pstack guide, chapter 2
What's on a card
Every skill is a file called SKILL.md. The first few lines are labels. The rest is plain instructions. Tap a highlighted line.
---name: principle-prove-it-worksdescription: "Apply after completing a task,
before declaring done. Verify against the real
thing, not a proxy or 'it compiles.'"disable-model-invocation: true---
# Prove It Works
Verify every task output by checking the
real thing directly.
**Why:** Unverified work has unknown
correctness.
## Script the check when you can
The strongest proof is a script that
re-runs the same comparison.Four kinds of cards in pstack
The head recipe and table of contents. You type it; it reads your request, picks the right playbook, and calls other skills as each step needs them.
Step-by-step recipes for each kind of job (fix a bug, build a feature, ship). They live inside the router and get picked for you. You never type them.
Routines for one job: explain code, get a second opinion, try ideas side by side, clean up writing. Type them yourself or let the router call them.
One-rule cards like "Prove it works." You steer with their names; the AI opens one when its moment comes and must say which rules shaped its decisions.
A card's journey
Here's exactly what happens to one skill, /how, from sitting in a folder to giving you an answer. Step through it.
Almost no card loads by itself
57 of the 58 skills are marked "manual only," so the AI can't see them in its list of skills. A card gets read in only two ways: you type its name, or another card points to it ("now use the how skill"). So pstack works like a chain, each card handing off to the next:
Just names and one-line descriptions. Manual-only skills are hidden from it.
Must-do rules, the list of principles, rules for helpers, and the menu of 23 playbooks.
Numbered steps, copied word for word into the AI's to-do list. A skipped step stays listed with its reason.
Loaded at the exact step that needs them, then followed.
Instructions for helper AIs. They load in each helper's own memory, so the lead's stays clean.
Which skill, when
Real work moves through six stages, and pstack has skills for each one. Pick a stage to see what happens there, which skills help, what you might type, and Lauren's advice for that stage.
All 34 skills you can type
The tool skills plus the router and setup. In Claude Code, put pstack: in front, like /pstack:how. Tap any skill for details.
The 23 playbooks
You never type these. Start with /pstack:poteto-mode, describe the job in your own words, and the router picks the matching playbook.
Leaving it running overnight
pstack can run long jobs while you're away, but only once you trust the loop. The guide says to give every unattended run four things:
In the morning, /show-me-your-work has logged every decision, and /ketchup gives you a quick catch-up.
The most important skill
Lauren's guide calls verification the foundation of everything she does with agents. Verification means the AI can check its own work on the real thing, so it keeps going until it has actually succeeded instead of guessing.
- The AI makes a change and says "should work."
- You open the app and try it.
- It doesn't work, so you explain what's wrong.
- Repeat. Your day goes to babysitting.
- The AI makes the problem happen in the real app.
- It makes a change and tries again.
- It repeats until the problem is truly gone.
- It hands you the proof: screenshots, output, video.
What /create-verification-skill builds for your project
A small tool the AI can run to open your app, click, type, take screenshots and read what's on screen, the way a person would.
Notes listing every feature and how to reach it. Lauren calls this "materialized memory": knowledge written down so every new AI starts with it.
How to start the app, load test data and log in as a test user, done the same way every time.
Keep it fresh. Apps change, so the map goes stale. /maintain-verification-skill updates it, and Lauren suggests running it at least once a day. Good remote controls also have clear help text, a safe "dry run" option for anything destructive, and error messages that tell the AI what to do instead.
Two sayings from the guide
"It compiles" is not evidence.pstack guide, chapter 6
Passing automatic checks doesn't prove a change is right. The Shipping playbook has an independent helper re-check each change before it's merged.
Going further: proof by math
/create-formal-verification-skill (in beta) sets up . Instead of trying a few cases, a checker tests every possible situation in scope against rules that must always be true. Lauren reports it has already found real bugs. /statechart turns those rules into a diagram people can read.
The memory game
The AI can only answer from what's in its memory. Tap items to put them in, then ask a question. This memory holds 50,000 tokens (a real one holds about 200,000) so you can fill it. Then overfill it and see what gets lost.
Pick a question to see what the AI can answer from its memory right now.
One lead, many helpers
An "agent" is just a separate AI conversation with its own memory. The lead, which is the chat you typed into, can start helpers for parts of the job. Helpers report back, and the lead decides what to keep. Lauren describes doing this well as supervising someone smarter than you.
Your chat. It owns the to-do list, makes every decision, reviews every change and writes the final answer.
A coding helper whose one rule is to read the router card in full first, so it follows the same handbook as the lead.
Helpers that can look but can't change anything. They're used to study code and to find problems.
Many helpers at once, each in its own copy of the project so they can't trip over each other.
A gleefully fanatical character that hunts down pointless code comments before review. It only reports.
Helpers that score other helpers' work against a checklist, ideally using a different AI model from the lead.
Six rules for helpers
These come from the router card, and most of pstack's reliability comes from them.
- The lead runs the plan. Helpers only do steps.Decisions stay in one place.
- The lead owns every helper's work.It reviews the change itself. A helper saying "done" isn't proof.
- A fresh helper for each new job.A retry goes to a new helper with the full brief, not an old one with a cluttered memory.
- Point to files, don't paste them.Helpers read what they need, and the lead's memory stays clean.
- The right model for the role.Fast models for routine work, the strongest for the hardest calls. A second opinion means a different model.
- Parallel only when the work splits cleanly.Separate files can be worked on side by side. Shared files go one at a time.
When does it ask you?
The router card is clear about this.
- Anything that can be undone
- Posting to team chat, updating tickets
- Running experiments to answer its own questions, instead of asking you
- Overwriting shared history (force-push)
- Deploying
- Deleting data
- Messaging customers
It's also allowed to say no. If an idea doesn't earn its place, the router card tells it to push back rather than agree by default.
Six team patterns
Every multi-helper skill has the same shape: set up the job, send out helpers, gather the results, and let one place decide.
The upside-down pyramid
How do you trust work an AI does faster than you can read it? Lauren's answer is a pyramid turned upside down: put the most effort where it pays off forever, and the least where it doesn't scale.
If you keep correcting the AI for the same mistake, correcting it again in chat won't stick. Fix the environment instead. /correct finds mistakes agents repeat and makes each one impossible, trying the strongest fix first: better structure, then automatic checks, then a test, and a written rule only as a last resort.
Lauren compares a fast-changing project to a bonsai tree growing in every direction. You don't fight the growth. You shape it with the right constraints, so it grows the way you want on its own.
Eight things that sound true
Most confusion about AI agents comes from a few reasonable-sounding assumptions. Tap a card to flip it.
The 24 principles
Each principle is its own small card, and the router keeps a one-line list of when each applies. The AI opens the full card when its moment comes, and its final answer has to name the rules that changed a decision.
Six levers, small to big
pstack is one person's style. Start small: watch where agents fail, and add a skill or a check only when the same failure shows up twice.
Run /pstack:setup-pstack. Pick a budget (how hard the AI thinks) and which model does which job. Smaller budgets use fewer tokens.
Three helpers review the conversation, a fourth sorts their ideas, and nothing changes until you approve. Edits go into your own copy, never the original.
Turns a correction you keep making into a permanent fix in the project itself.
It reads your past chats, finds your habits, asks which ones really fit you, and drafts your personal version of poteto-mode.
Ask /pstack:poteto-mode write a skill for …. Keep only sentences that change a decision, and test a skill change before you trust it.
Make your own copy of the pstack project (a fork), edit it, and share improvements back. Lauren welcomes contributions.
Two rules for safe edits. Never edit the installed copy of pstack, because updates overwrite it. Put your own skills in your own skills folder instead. Also, a new card does nothing until something points to it: a new playbook needs its file and one line in the router's menu.
Where everything lives
The real layout of the pstack folder. Tap any item to see what it does and what you'd change there.
Templates
Models and Claude Code
small (medium effort, cheapest) · medium (high) · large (extra high, the default) · unlimited (maximum). Your choices are saved to one settings file that every skill checks first.
Commands get a pstack: prefix, and Cursor-only features are skipped with a one-line note. pstack's default team mixes Claude and Grok models, but Claude Code only runs Claude, so "second opinion" skills work best when you assign two different Claude models (for example Opus and Sonnet).
/plugin marketplace add cursor/plugins
/plugin install pstack@cursor-pluginsInstall in Claude Code. In Cursor: /add-plugin pstack.
Eight quick questions
If you can answer these, you understand how pstack works. Every answer explains itself.