pstack for Claude Code

This is a port, not original work. pstack was designed and written by poteto and published in cursor/plugins. Every skill, playbook, and principle here is hers. This repository only translates them to run on Claude Code. Not affiliated with poteto or Cursor.

If you want to go fast, go deep first. That is pstack's whole argument. It is a set of skills for writing less code of higher quality, and for parallelizing agents you can actually trust. The goal is not to maximize lines of code. It is the opposite.

What changed in the port is everything that names a tool, a model, or a config path. The Task tool became the Agent tool. Cursor's model roster became Claude's. Cursor's hosted Automations became GitHub Actions. Read PORTING.md for the full translation table, what got better, and the one thing that genuinely does not survive the move.

If you find a bug here, check whether it exists upstream before reporting it. A porting mistake is mine. A design question is poteto's.

Install

Clone, then run the installer for your shell.

macOS, Linux, Git Bash

git clone https://github.com/negativeblueprint/pstack-claude.git && cd pstack-claude && ./install.sh

Windows PowerShell

git clone https://github.com/negativeblueprint/pstack-claude.git; cd pstack-claude; .\install.ps1

This copies 44 skills into ~/.claude/skills/ and 2 agents into ~/.claude/agents/. Restart Claude Code afterward, since skills are discovered at session start.

The installer is safe to re-run. It compares content, so a second run reports everything already current and changes nothing. If you already have a skill with one of our names (how, why, teach, swarm and bro are the plausible collisions), it refuses rather than overwriting, and tells you which ones. Pass --force (bash) or -Force (PowerShell) once you've backed those up.

To install somewhere else, set CLAUDE_HOME.

To check the installer on your own machine before trusting it with your config, ./verify-install.sh runs the whole thing against a temp directory and never touches ~/.claude.

Uninstall by deleting what it copied.

cd pstack-claude && ls skills | xargs -I{} rm -rf ~/.claude/skills/{} && rm -f ~/.claude/agents/poteto-agent.md ~/.claude/agents/comment-sicko.md

Installing as a plugin instead

The repo is also a self-contained plugin marketplace, which gets you /pstack: namespacing and one-command removal. It needs Claude Code's interactive plugin panel, so it works in the terminal but not in every client.

/plugin marketplace add ./pstack-claude

/plugin install pstack@pstack-claude

Pick a scope when the details view opens. If the summary says Run /reload-plugins to activate., run that.

Use one method or the other, not both. Two copies of the same skill name is exactly the collision the installer is built to avoid.

Get started

Two steps.

1. Run /setup-pstack and choose which models fill which roles.
2. Use /poteto-mode whenever you're doing anything that needs rigor.

The guide walks through a first real task, from setup and prompting through verification and overnight runs.

Everything else is situational. /poteto-mode routes to the other skills as its steps need them.

The main entry point

/poteto-mode reads your request, matches it to a playbook, and runs the other skills as the steps fire.

/poteto-mode this pr has a subtle bug where the scroll drifts every 750ms even when idle. repro
first, then fix and verify.

When invoked it opens a todo list whose first item is reading the inline principles index, copies the matched playbook's steps in verbatim, routes to the other skills as steps fire, and writes an unslopped reply framed for both the consumer and the maintainer.

Twenty-two playbooks ship with it, covering investigation, bug fixes, perf, hillclimbing, runtime and trace forensics, features, refactors, prototypes, visual parity, skill authoring, evals, babysitting a PR to green, shipping a stack, autonomous runs, orchestration, autopilot, session pickup, pausing safely, multi-phase plans, and worktree cleanup. They live in skills/poteto-mode/playbooks/.

/poteto-mode pairs well with Claude Code's built-in /loop, which is how you get long unattended runs without dropping rigor.

Skills

44 skills. /poteto-mode runs most of them for you. Reach for one directly when you want just that.

skill	use it when
/poteto-mode	default entry point for any non-trivial task.
/how	you want a walkthrough of how a subsystem works.
/why	you want to know why something was built this way, from source control, tickets, docs, chat, and observability.
/recall	you're resuming work and want your recent context on a topic rebuilt into a current-state brief.
/blast-radius	a small-looking change, and you want to know what else it could break, proven by running code.
/architect	you're about to write code crossing a function boundary and want types and module shape settled first.
/arena	you want N parallel attempts at the same thing, then the best parts of each.
/swarm	you want N parallel workers across slices or races, then one aggregated report.
/interrogate	you have a diff and want several reviewers trying to break it.
/reflect	you want the session's own transcript reviewed for lessons worth encoding.
/figure-it-out	no bundled playbook fits and you want a rigorous one designed for the task.
/unslop	any prose surface. cuts AI tells.
/no-comments	you want the narrating comments gone before review.
/technical-writing	docs, RFCs, readmes, PR descriptions, commit messages.
/teach	you want a body of work explained so a person actually understands it.
/tdd	you explicitly want a failing test first, or the bug has an obvious regression test.
/show-me-your-work	long or unattended work that needs an auditable decision trail.
/create-verification-skill	your project has no way for an agent to drive the real app and prove a change works.
/maintain-verification-skill	that skill has drifted from the app.
/automate-me	you want your own working style mined out of your history into a personal mode skill.
/setup-pstack	you want to change which model fills which role.
/bro	the last message was jargon. say it like a human.

Plus /typescript-best-practices, and twenty principle-* skills that the mode's principles index navigates into.

Principles

The principles are the load-bearing part. Each is a leaf skill that /poteto-mode navigates into when it applies, grouped as core, architecture, verification, delegation, and meta. They cover things like biasing to deletion, modeling the domain in a structure instead of scattered conditionals, making illegal states unrepresentable, proving work against the real artifact rather than a proxy, tracing symptoms to root causes, guarding the context window, and not blocking on the human for reversible work.

The index lives inline in skills/poteto-mode/SKILL.md.

Agents

poteto-agent runs the full style end to end. Spawn it with subagent_type: "poteto-agent". It reads poteto-mode in full, principles index included, before doing any work. Substituting general-purpose skips that read and drifts.

comment-sicko is a read-only comment reviewer. Usually reached through /no-comments rather than directly.

Automations

automations/benny/ is a separate pack, not part of the installed plugin. It triages Slack issue reports and reproduces confirmed bugs. Upstream it ran on Cursor's hosted Automations product; here it runs as GitHub Actions workflows invoking Claude Code headlessly. Read automations/benny/FOR_AGENTS.md to install it into a target repository.

Credits

pstack is by poteto. The skills, the playbooks, the principles, the twenty-two-playbook router, the whole idea of encoding engineering judgment as leaf skills an agent navigates into, all of it is her design. Read the original first if you want the source of the ideas.

This port exists because the ideas are good and the format is portable. Anthropic's SKILL.md spec is what both editors read, so most of pstack moved across unchanged. That is a nice property of an open format, and it is worth saying out loud that the port was easy because someone else did the hard part.

Upstream is MIT licensed and that license is preserved in LICENSE. The original README is kept verbatim at README.upstream.md so you can compare. Fork it, improve it, make it yours.