Open‑Source ‘pstack‑claude’ Brings Cursor‑Based Agent Workflows to Claude Code, Copilot and More

Overview

The community‑driven pstack‑claude repository (github.com/michael-denyer/pstack-claude) publishes a port of Lauren Tan’s *pstack* skill stack for a growing list of AI‑coding agents. The project translates the original Cursor‑style primitives into formats that Claude Code, GitHub Copilot, Gemini, Muse, Prime Agent and other harnesses can consume.

Key Features

• Unified skill library – One set of tools (e.g., sub‑agent, question, wake‑up, /loop) works across all supported agents.
• Formal‑verification plug‑in – `/create-formal-verification-skill` automatically selects the riskiest property in your code and runs the appropriate verifier (TLA+, Lean, Kani for Rust).
• Automatic routing – A *SessionStart* hook injects an instruction that directs the agent to *poteto‑mode* so the correct workflow is chosen without manual prompting.
• Fine‑grained control – `setup-pstack` lets users toggle auto‑routing, change default models, or adjust reasoning effort per role.
• No telemetry – All scripts run locally; no data is sent to a central server.

Supported Harnesses

Installation & Updates

1. Clone the repo and run the provided for‑loop script to link the skill files into the appropriate directories.
2. For Claude Code, enable auto‑update via the *Marketplaces* tab → *pstack‑claude* → Enable auto‑update; otherwise run `claude plugin update pstack@pstack-claude` for a manual bump.
3. Copilot users should start a new session after installation; the first skill that needs a model will invoke `setup-pstack`.
4. Keep the code current with `git pull` followed by the linking script.

Practical Implications

• Developers get a consistent, opinionated workflow for writing, testing and formally verifying code across multiple AI assistants.
• Teams can embed the same verification standards (TLA+, Lean, Kani) without writing custom plug‑ins for each agent.
• Security‑focused projects benefit from the local‑only execution model, which eliminates hidden telemetry.
• Tool vendors may adopt the open‑source skill definitions as a baseline for their own agent extensions.

How to Verify the Claims

• Review the `README.md` and `tools/forks.json` in the GitHub repository – they list the exact skill set and policy forks.
• Check the `LICENSE*` files to confirm the MIT licensing and attribution to Lauren Tan, Cursor and Michael Denyer.
• Run `pstack‑claude --help` after installation; the output should list the `/skill:` commands and the `/create-formal-verification-skill` entry.
• Test the routing hook by starting a Claude Code session and issuing `/poteto-mode`; the agent should respond with the appropriate workflow.

Licensing & Contributions

The code is released under the MIT license (© 2026 Michael Denyer) with original pstack attribution (© 2026 Lauren Tan). Contributors are encouraged to open pull requests, file bug reports, or improve documentation as described in `CONTRIBUTING.md`. Security issues should be reported privately per `SECURITY.md`.

--- *The article is based solely on the public information available in the GitHub repository linked above.*