Special · pstack

Wake up to thousands of merged PRs: what 'pstack' is and why X won't stop talking about it

"🫡 time to install pstack." Thariq, who worked on Claude Code for over a year and left Anthropic on Oct. 24, posted that the next day to 840 likes, right after his farewell post drew 30,500. pstack is a skill pack for coding agents built by Lauren Tan (@poteto), who has worked at Netflix, on Meta's React core team and at Cursor. Her claim of roughly 2,500 agent-handled PRs a month has made it a fixture of developer X. We went to the sources to explain what it is and why it matters.

1. cursor/plugins · GitHub ↗

   What pstack is: a playbook for how agents should work

   pstack is an MIT-licensed skill pack in Cursor's official plugin repo (cursor/plugins). It holds 23 Markdown playbooks, 24 principles, 30-odd skills and two subagents, laying out the steps an agent should follow for bug fixes, perf work, refactors and features. The entry point is /poteto-mode: put it in front of a prompt and it picks the right playbook and calls other skills.

2. cursor/plugins · GitHub ↗

   Why it's spreading: no more "fixed it" when it isn't

   A common agent failure: edit a few files, run tests, report success, and the bug is still there. pstack's bug-fix playbook makes the agent reproduce the failure in the running app before touching code, find the root cause, make the smallest fix and re-run the original failing interaction. The README says "if you want to go fast, go deep first" and "the goal is not to maximize loc, in fact it's the opposite."

3. @MaiYangAI · X ↗

   Merging while you sleep, with separate verifiers

   The autopilot playbooks drew the most attention. One owner agent carries each PR to merge, but its own verdict doesn't count: separate verifier agents must sign off on the exact patch, and CI must pass on a freshly rebased head. Tan told Matt Pocock she hands work to agents, goes to sleep and checks in the morning.

4. MixedLauren Tan · X ↗

   "2,500 PRs a month": a self-reported number

   Tan has cited roughly 2,000 PRs a month (late August) and about 2,500 (a September talk). The figures aren't independently verified, and she says "they're not obviously like 2,500 features": much of it is gardening, such as restructuring and deleting code. Full autopilot is "quite token intensive."

5. @cu30rry_ · X ↗

   It works in Claude Code too

   Built for Cursor, pstack gained official Claude Code support on Oct. 17: /plugin marketplace add cursor/plugins, then /plugin install pstack@cursor-plugins, with commands prefixed like /pstack:poteto-mode. The post announcing it drew 2,100 likes. An unofficial port for Codex, Copilot and OpenCode, pstack-claude, was reported at over 2,000 GitHub stars.

6. Positive@DanielLockyer · X ↗

   What users say: "deploys 30% faster"

   Developer Daniel Lockyer said pstack and Opus 5.5 cut his production deploy time by about 30%. Developer dotta called it "a goldmine of prompt engineering," praising Tan's sample prompt for packing a goal, a checkable finish condition, the evidence to show and constraints. Dmitriy Leybel, who had Opus 5.5 write a pstack deep dive, called it "a masterclass in agentic engineering."

7. Lauren Tan · X ↗

   What's next: statecharts and formal verification

   Tan recently showed agents drawing statecharts of code, then using a pstack-made formal verification skill to hunt bugs with tools like TLA+; those posts each passed 5,000 likes. On Oct. 11 she said pstack got 10% smaller with no loss in behavior, and she expects most skills to compress into short principles as models improve.

8. MixedLauren Tan · developer.tenten.co ↗

   Limits: not plug-and-play

   Tan says the setup took months and is "very hard to get to this point." Multi-agent verification burns usage quickly, so one implementation plus one focused review is often enough for routine work. She has no answer yet for domains like finance or healthcare where changes can't be verified or reverted automatically. Irreversible actions still go to a human.