brain

pstack

Properties7
statusseedling
typeconcept
created2026-10-03
updated2026-10-05
tagspstack, agents, cursor, verification, delegation, ai-engineering
aliasespstack, poteto-mode
public

4 min read

pstack

A Cursor plugin by Lauren Tan (poteto) — Cursor engineer, React core team, ex-Meta/Netflix — that packages her daily work procedures and judgment standards as Playbooks, Principles, and Skills (as of v0.15.5, 2026-09-24: 23 Playbooks, 23 Principles, 24 Skills; lives in cursor/plugins). The headline number: 2,500 PRs shipped to production in the single month of August 2026 (announcement).

The framing that matters more than the number — the book’s central question: “Under what conditions can an agent keep doing high-quality work without a human watching?” Getting code written is no longer hard; building a system that lets you trust it is. Otherwise “however many agents you run side by side, the human who checks their work stays the bottleneck.”

Delegation, per the book, means handing the agent three things:

  1. What counts as done — the finish condition as a runnable pass/fail
  2. How to proceed — investigation → reproduction → fix → checking, same quality every time
  3. How to verify — the agent itself runs the real app/artifact and leaves evidence

If the third is missing, a human is still the checker — which is why poteto calls verification the most important foundation, and why the example request is shaped like /poteto-mode the export writes duplicate rows when a retry lands mid-run. repro first, then fix and verify. — symptom + definition-of-done in two sentences.

Kinship map (this is the vault’s AI-engineering cluster converging): the verification-first stance is Agent tests and trajectory grading; parallel-trusted-agents is the AI engineering control plane; “procedures as reviewable artifacts in the codebase” is the Golden repo / Markdown Slideshows as-code family; and it operationalizes Operating principles (“engineer systems so everyone can ship safely” — here the ‘everyone’ is agents).

Resources

Related: AI engineering · AI engineering stretch goals · Agent tests

The guide (Part 1, Aug 2026) — x-2094457600259842065

lauren’s own Complete Guide to pstack Pt. 1 (clippings/x-2094457600259842065.md, 2-tweet thread with embedded article): shipped on Grok @Bot — “2,000 PRs a month” reframed as gardener/maintainer role: “Being Grok @Bot’s gardener and maintainer is something I was only able to achieve through pstack… you can only do that when the foundations can keep up with the number of contributions.” The window-of-opportunity move: refactor the codebase into strong foundations while it’s being built, with no downtime, as hundreds of PRs/day land.

Concrete prompt patterns from the guide (the /poteto-mode grammar):

  • /poteto-mode create a plan to migrate our entire UI library to StyleX. break the migration into small, verifiable PRs… i want the final result to be 100% identical compared to the original — bugs included (characterization-testing discipline as a prompt)
  • /poteto-mode repro this with /control-app. if it repros on main, fix it and show me a video as proof (agent-run verification with evidence)
  • /poteto-mode do it (thread already has context — trust the contract)

And the anti-pattern named: “abstract plans only give you the illusion of progress” — Plan Mode as self-persuasion; pstack’s replacement is “thorough investigation, empirical evidence, and rigorous verification. When you plan this way, engineering with agents stops feeling like a gamble.”

Part 2/3 continuation links live in the clipping.

The Bend connection (2/2 of the 2500-PRs thread)

The thread’s second tweet points at the deeper bet: bend (OSS) (HigherOrderCO, 23.4k★) — “very bullish on Bend, this is exactly what I’ve been trying to do in other languages, but you can only get so far with weak type systems and lint rules. When you have a codebase that can formally verify itself, you can ship at an incredible pace. Code review is solved.” The pstack verification discipline (lints, evals, evidence) as a manual practice meets its theoretical endpoint: formal verification in the language itself. Same family as Datadog Temper’s “the spec is both proved and executed” — the convergence of the agent-era stack toward verifiable artifacts. The 2,500-PRs video itself (1.5M views): 3442x2160 MP4, tweet-linked; no transcript exists — the zenn book, guide Pt.1, Pocock LIVE, and Perfology interview remain the textual sources.

Graph View

Backlinks