Pstack Skill

The Problem

AI agents write code fast — and often slop. Without a method, the agent guesses at implementations without understanding the system, ships without verifying real behavior, stacks untested changes, and stops to ask questions a quick experiment would answer. Throughput without quality is not a goal: you want less code, but better.

Lauren Tan’s (@poteto) pstack solves this — but it was a Cursor-only plugin tied to marketplace installs, cloud agents, Graphite, and vendor model slugs.

The Solution

Pstack Skill is the self-contained port of the full pstack stack into a single skills.sh-compatible skill. One orchestrator that:

• reads your task and picks one of 23 playbooks (bug fix, feature, refactoring, perf, investigation, prototype, babysit, shipping, autonomous run, orchestrate…);
• copies the playbook’s steps verbatim onto a todolist before reasoning about the task;
• routes to 21 bundled procedures as steps fire (how, why, architect, arena, swarm, interrogate, unslop, technical-writing, show-me-your-work…);
• applies and cites 21 engineering principles, each citation traceable to a real decision;
• verifies on the real artifact — runs the feature like a user would, never trusts “it compiles”;
• is sticky: once invoked it stays on across turns, applying itself when rigor is needed and staying out of the way otherwise.

Based on pstack by Lauren Tan, MIT License, informed by the open-pstack port. Cursor-specific mechanics were translated to platform-agnostic equivalents.

When to Use

• Nontrivial code changes, architecture decisions, or “are we sure?”
• Scientific debugging: reproduce → root cause → failing-test-first fix
• Adversarial multi-model review before shipping contested designs
• PRs that need to reach merge-ready (conflicts, review threads, flaky CI)
• Long or unattended work (“going to bed, finish this”) with an auditable trail
• Multi-day programs with dozens of subagents under one coordinator

Installation

npx skills add https://github.com/fabricioctelles/skills -s pstack-skill

Or manually: copy the skills/pstack-skill/ folder to .claude/skills/, .cursor/skills/, .kiro/skills/, or your agent’s equivalent directory.

How It Works

task in plain language
        |
        v
+---------------------------+
| orchestrator (SKILL.md)   |  reads 21 principles
| picks the playbook        |  opens todolist
+---------------------------+
        |
        v
+---------------------------+     +--------------------+
| playbook (23 options)     | --> | procedures (21)    |
| steps verbatim on list    |     | how/arena/swarm/   |
+---------------------------+     | interrogate/unslop |
        |                         +--------------------+
        v
+---------------------------+
| verify on real artifact   |  real surface, never a proxy
| decision trail (TSV)      |  auditable after the fact
+---------------------------+
        |
        v
    merge-ready PR

Model Roles

Delegations never name vendors. Four role slugs with capability contracts, bound once in .agents/pstack-models.md:

Role	Contract	Typical work
worker	fast, cheap instruction-following	trivial edits, explorers, swarm
builder	strongest instruction-follower, long context	specified implementation
judge	deepest reasoning + calibrated prose	synthesis, reviews, cross-judge
peer	strong reasoner from a different family than judge	panel diversity

One model available? All four collapse to it and panels become sequential independent passes on fresh context — no verification gate gets skipped.

The Playbooks

Scenario	Playbooks
Understand before touching	Investigation, Runtime forensics, Trace forensics
Build	Feature, Prototype, Refactoring, Visual parity, Multi-phase plan
Fix scientifically	Bug fix, Perf issue, Hillclimb
Ship PRs	Babysit, Shipping, Opening a PR, Autopilot-full, Autopilot-stack
Auditable autonomy	Autonomous run, Orchestrate, Session pickup, Pause safely
Agent tooling	Authoring-a-skill, Eval, Worktree cleanup

Highlights

• Repro-first: a bug you cannot reproduce is a bug you cannot prove fixed
• Green ≠ safe: shipping only after an independent per-PR verdict, landing only the contiguous verified run from the root
• Never block on the human: reversible work proceeds; only ask what no experiment can settle
• Decision trail: append-only TSV with per-row evidence, cross-reviewed by a different model before handback

Compatibility

Works on Claude Code, Codex, Cursor, opencode, Kiro, and any agent that reads skills.sh-format SKILL.md files — with one or many models, with or without alternative CLIs.