# How I Use Cursor (Lauren Tan / poteto)

Pinned X Article in which Cursor engineer Lauren Tan (@poteto) open-sources **pstack** — her daily-driver skills/playbooks — and argues that agent parallelism only works after you can trust one agent to finish and verify a problem. Written May 2026 while she was on Cursor 3’s Agent Window; by ingest (Aug 2026) her bio is Grok Bot + Cursor at SpaceXAI.

Tags: source, x-article, cursor, pstack, poteto-mode, verification, agent-orchestration, grok-bot

## Who / what

| Field | Value |
|---|---|
| Author | lauren / Lauren Tan (@poteto) — React compiler core team; prev Meta, Netflix EM; ~35.7k followers at ingest |
| Form | X Article + pinned announcement post (body is the article card) |
| Title | How I Use Cursor |
| Created | 2026-05-25T18:15:06Z (2:15 PM ET) |
| Engagement (~2026-08-17) | 1424 likes · 2775 bookmarks · ~350k impressions · 130 reposts · 57 replies · 48 quotes |
| Post | https://x.com/poteto/status/2058975157503570132 |
| Article | https://x.com/i/article/2057201109002059776 |
| Site | https://no.lol |

Primary body: tool.lu cache → raw/`x-poteto-2058975157503570132-2026-08-17/article.md` (X API had title only; see notes).

## Bottom line (for Josh)

1. **Thesis:** “If you want to go fast, go deep first.” Naive multi-agent fan-out writes slop faster. Parallelism is a trust problem, not a process-count problem.
2. **pstack / `/poteto-mode`:** her Cursor plugin. Higher-order skill picks a playbook (debug, perf, feature, prototype, overnight land, visual parity, …) and routes other skills. Goal is **least code, most impact**, not LOC.
3. **Verification is the bottleneck.** She will not automate a factory until an agent can own a bug end-to-end including repro + proof. Same spirit as [[grok-4-6-field-guide]]’s “what done means,” different artifact (playbooks vs one-liner).
4. **Benny:** Slack-triggered Cursor Cloud Agent that triages (images/video, git, Slack, Notion), then a second Benny `/orchestrate`s computer-use repro → fix → traces/video → PR. WIP in May; pstack later ships a dormant `automations/benny/` pack.
5. **Affiliation drift:** article is Cursor-employee, Cursor 3 Agent Window, Cursor automations. Do not treat it as a [[grok-bot]] tutorial. Bio at ingest says she now works on Grok Bot + Cursor at SpaceXAI — useful lineage, not the source’s subject.

## What the article actually says

### Why she switched (Claude Code → Cursor 3)

Never used Cursor before the interview. At Meta she paid $200/mo for Claude Code, loved the CLI, and was building her own orchestrator on top — which she later reads as Boris Cherny’s “latent demand”: a hackable CLI makes the *human* the orchestrator, so every team wraps it. Onsite (pre-Cursor 3, Editor Window) three things landed: native multi-model (Opus frontend / Codex systems + mixed subagents), **fast compaction** (cc compaction felt slow and “super dumb” after), and GUI affordances (in-editor browser + Design Mode). Joined end of March 2026; daily driver is Cursor 3 Agent Window.

### Agents as amnesiac new hires

EM frame: onboard for codebase *and* how work gets done. Agents don’t remember or learn unless you attach rules, skills, tools, long-term memory. Failure modes become playbooks. Without rigor they sycophantically emit volume.

Embedded Apr 26 note: value of many agents is **depth** (best-of-N on one problem), not breadth.

### pstack (article-era surface)

Open-sourced as the skills she uses to build Cursor. Install: `/add-plugin pstack` · https://cursor.com/marketplace/cursor/pstack. Claim: among the most-used internal skills; **9k uses that week** on the company leaderboard.

`/poteto-mode` is the heart. Example prompts in the article (scroll-drift bug, virtualized-list CPU trace, feature-flag, dual markdown-renderer prototypes, OSS-the-plugin, overnight CI-flake land, visual-parity from a screenshot).

On-demand skills named in May: `/how`, `/why` (parallel MCP evidence: git, issues, docs, chat, infra, errors, warehouse), `/architect`, `/arena`, `/interrogate`, `/tdd`, `/unslop`, `/reflect`, `/figure-it-out`, `/show-me-your-work` (TSV decision trail), `/automate-me` (mine transcripts → your own `-mode`). Works in other agentic tools; designed for multi-model Cursor.

### Benny / “software factory”

Cursor automations = scheduled or Slack-event cloud agents. Benny (WIP): triage employee RC feedback → ticket (code + git regression + Slack + Notion “bug vs designed”) → second Benny via `/orchestrate`.

Embedded May 8 `/orchestrate` note (Cursor SDK, recursive spawn): internal-skills autoresearch **−20% tokens / better evals**; internal backend cold start **−80%**.

Repro path: Cloud Agents run Cursor in the cloud (desktop click/keyboard; CDP-or-equivalent control skills). Consistent repro → fix; perf → before/after CPU + heap; workers verify against the ticket, shoot before/after video, open the PR. Code-review scale called out as the next unsolved.

Closer: do not parallelize agents you don’t trust — wasted tokens + slop.

## pstack as of ingest (2026-08-17) — later than the article

Verified on `cursor/plugins` `pstack/` (plugin.json **0.14.1**, MIT, author Lauren Tan). Grew past the May list:

- **22 playbooks** including hillclimb, babysit, shipping, orchestrate, autopilot-full/stack, session pickup, pause safely, worktree cleanup.
- Extra skills: `/setup-pstack` (per-role models — the answer to the May 26 token-cost reply), `/recall`, `/blast-radius`, `/swarm`, `/teach`, `/no-comments`, `/bro`, `/technical-writing`, create/maintain-verification-skill.
- 21 principle skills (laziness-protocol, prove-it-works, never-block-on-the-human, …).
- Subagents: `poteto-agent`, Comment Sicko.
- README default split: specified code → sol; fast mechanical → grok; prose/judgment → fable (panel fable / sol / grok / opus 5).
- Dormant Benny pack: `pstack/automations/benny/` (not slash skills).
- Sibling: `cursor-team-kit` for `/deslop`, `control-cli`, `control-ui`.

Author-linked examples the day after publish: https://github.com/cursor/plugins/blob/main/pstack/README.md#examples

## Author replies (2026-05-26)

- pstack **complements** “brainmaxxing,” not a replacement.
- **Composer 2.5** for speed/efficiency; frontier for harder work (quota pushback).
- Thinks slash-command memorization can go away.
- Parked the “downgrade every skill’s model” request (`/setup-pstack` exists later).

Later community: some prefer pstack to Superpowers in Cursor; Aug 2026 still asking whether Benny-style automations move to [[grok-bot]] (unanswered in the sampled thread).

## Open questions

1. Exact IDs of the embedded Apr 26 / May 8 posts (search 429; not recovered).
2. Did Benny-style Slack automations move onto Grok Bot after her SpaceXAI move? Article does not say; 2026-08-14 reply asks.
3. How much of the May playbook list maps 1:1 onto the Aug 22-playbook table vs renamed.
4. Internal “9k uses this week” — one-week snapshot, not a durable stat.

## Related

- [[grok-4-6-field-guide]] — other Cursor engineer X Article; verification-loop cousin
- [[grok-bot]] / [[grok-bot-community-usage]] — later affiliation, not this article’s product
- [[hermes-agent]] — named in a reply as a possible poteto-mode host
- raw/`x-poteto-2058975157503570132-2026-08-17/`

## Sources

1. **Primary post/meta:** X API — raw/`…/story.json`, `posts.md`
2. **Article body (secondary cache):** tool.lu preview of the X Article — raw/`…/article.md`
3. pstack marketplace + GitHub README / plugin.json 0.14.1 / PR #73
4. Author replies listed in raw/`…/posts.md`