Frontend Design
Use Claude Code to create production-ready interfaces that look professional, not AI-generated. React, Tailwind, and clean design automatically.
Routes non-trivial engineering work through poteto-mode playbooks with verification, multi-model review, and a review-ready PR.
Grok Bot, SpaceX AI
Routes non-trivial engineering work through poteto-mode playbooks with verification, multi-model review, and a review-ready PR.
What it tells your AI to do
Someone else wrote this — read the source, verify it does what it says, and never share a password or API key with it.
Use poteto-mode: add saved filters to search, learn how search works today, keep the design simple, verify in the real app, and open a review-ready PR. Create a verification skill if the project lacks one.
A skill is instructions your AI follows. Read them before you install — and never paste in a key or password one asks for.
# pstack You follow Lauren Tan's pstack approach for rigorous engineering with AI coding agents (Claude Code, Codex; original also ships for Cursor). Prefer verification over more markdown instructions. ## When to use Use for non-trivial engineering tasks: features, refactors, investigations, architecture decisions, and PR babysitting. Do not use for one-line typos or pure chat. ## Core loop (poteto-mode) 1. State the goal in plain language. Do not list every skill you want to run. 2. Learn how the current system works before changing it (`how` / investigation). 3. Settle types and data structures before crossing boundaries (`architect`) when the design matters. 4. Implement the smallest complete change. Prefer simple over extra machinery. 5. Prove real behavior in the running app or with a project verification skill—not "tests passed" alone. 6. Challenge important decisions with multi-model review (`interrogate` / `arena`) when the stakes are high. 7. Carry the work through review, CI, and a ready-to-merge PR when asked (`babysit`). 8. After hard tasks, capture lessons that improve the next run (`reflect`). ## Useful skills to invoke by name - `how` — explain how a subsystem works - `why` — evidence for why it was built that way - `architect` — settle design before crossing boundaries - `arena` — several complete attempts, keep the best parts - `interrogate` — adversarial multi-model review - `create-verification-skill` / `maintain-verification-skill` — repeatable proof of real behavior - `babysit` — CI and review comments until merge-ready - `reflect` — improve skills after a hard task - `tdd`, `unslop`, `figure-it-out`, `show-me-your-work` — when the task matches ## Install pointers (do not invent alternate installers) - Cursor: Lauren's original plugin at `github.com/cursor/plugins` → `pstack` - Claude Code / Codex community ports: e.g. `ericlitman/open-pstack` (`/plugin marketplace add` then `/plugin install pstack@open-pstack`) or maintained forks such as `michael-denyer/pstack-claude` - Run setup (`/pstack:setup-pstack` or ask Codex to use `pstack:setup-pstack`) so model roles match what the user can actually run ## Rules - Verification first. Build or maintain a verification skill instead of piling on instructions. - Explicit user instructions beat auto-routing. - Do not claim a listing is live or publish anything. - Do not paste secrets. Do not download untrusted installers. ## Good output A bounded plan, evidence of how the system works today, the smallest change that meets the goal, real-behavior checks, and a review-ready PR (or a clear branch/commit report). Adapted from @poteto's approach: https://x.com/poteto/status/2094457600259842065 — upstream https://github.com/cursor/plugins/tree/main/pstack
No comments yet — be the first.
Use Claude Code to create production-ready interfaces that look professional, not AI-generated. React, Tailwind, and clean design automatically.
Use Claude Code to test web applications and find bugs before users do. Automated browser testing with plain English commands.
Use Claude Code to build MCP servers that connect Claude to your APIs, databases, and tools. Working server with auth and error handling in minutes.