Lauren Tan's pstack engineering skills, deeply ported to Devin.
Install once — your Devin sessions work in her style: deliberate sub-agents, unslopped prose, verified claims.
Unofficial community port — not affiliated with Cognition or Cursor.
pstack is Lauren Tan (@poteto)'s public set of agent skills for Cursor — 64 skills encoding how she works: fan-out sub-agents, adversarial review, slop removal, "prove it works" verification, and a set of named engineering principles.
devin-pstack is not a find-and-replace port. Every mechanism that exists only in Cursor has been rebuilt on the real Devin equivalent — child sessions, agent modes, reminders, automations, Devin Review, and the git_* tool family. Where Devin has no counterpart, the skill was redesigned rather than stubbed; where Devin has a better counterpart, the port takes advantage of it.
Devin gives these skills capabilities they never had on Cursor:
| Capability | On Devin | On Cursor (upstream) |
|---|---|---|
| Sub-agent isolation | Each devin_session_create child gets its own VM and checkout |
Shared workspace, manual worktree management |
| Cross-platform verification | fleet playbook fans the same change out to Windows/macOS/Linux children |
Single-machine only |
| PR watching | git_take_over_pr streams review comments + CI events live |
watch-pr polling scripts |
| Review triage | Devin Review findings feed a structured triage → end_triage loop |
Bugbot scraping heuristics |
| Orchestration | run_workflow journals every agent call and resumes after crashes |
Linear prompt chains |
| Long-running work | devin_reminder_manage RRULE schedules survive session sleep |
/loop timer prompts |
| Automations | devin_automation_manage — schedules, webhooks, Slack/GitHub/Linear/Jira triggers with approval cards |
Manual routines |
See docs/harness/devin.md for the full adaptation table.
In any Devin session, say:
install the plugin from github.com/Till3005/devin-pstack
Then once, in a session:
/setup-pstack
/setup-pstack detects which Devin agent modes your account has and writes your role→mode mapping to ~/.devin/rules/pstack-models.md. After that, start work with /poteto-mode.
| You say | What happens |
|---|---|
/poteto-mode ship this |
Runs the shipping playbook: draft-free PR, CI watch, review triage |
/swarm refactor the auth layer |
Fans N parallel workers out, drains them, returns one report |
/interrogate this design |
Multi-angle adversarial review before you commit to it |
/deslop |
Strips agent slop (dead comments, hedged prose, churn) from the diff |
/babysit PR 123 |
Takes over the PR: conflicts → threads → CI, in that order |
/ketchup |
Plain-words recap of what changed since you last looked |
64 skills total — 61 core skills in skills/, plus 3 bundled in the benny automation pack — grouped by what they do for you:
Orchestration & fan-out — run many agents, one answer
| Skill | What it does |
|---|---|
poteto-mode |
The flagship: her agent style — concise, deliberate sub-agents, verified claims. Playbook runner. |
swarm |
Fan N workers at one task, drain, single report |
arena |
N parallel candidates, pick a base, graft the best of the losers |
fleet (new) |
Cross-platform verification: same change tested on Windows/macOS/Linux child sessions |
interrogate |
Adversarial multi-angle review of a plan or diff |
automate-me |
Turns your repeated instructions into a durable -mode skill |
poteto-help |
Guide to picking the right skill, playbook, or principle |
setup-pstack |
One-time role→agent-mode mapping |
Playbooks live under skills/poteto-mode/playbooks/: shipping, babysit, autopilot-full, opening-a-pr, bug-fix, fleet, autonomous-run, and more.
Verification & review — prove it, don't claim it
| Skill | What it does |
|---|---|
control-ui |
Drive and verify a real UI — browser, Electron, web app — as a user would |
control-cli |
Same for CLIs and TUIs: flags, exit codes, output |
deslop |
Remove agent slop from a diff before it ships |
unslop |
Cut AI tells from any prose |
correct |
Find mistakes agents keep repeating; make each structurally impossible |
blast-radius |
What could this change break beyond the diff — with proof |
benchmark-checklist |
Vet a perf measurement before trusting it |
show-me-your-work |
Reviewable decision trail (TSV log) for unattended runs |
no-comments |
Spawn Comment Sicko; fix accepted findings; encode real constraints |
create-verification-skill / maintain-verification-skill |
Generate & keep honest a project-local verification skill |
create-formal-verification-skill / maintain-formal-verification-skill |
Same, for formal methods where a bug costs the most |
statechart |
Check a TLA+ model of a stateful system |
tdd |
Disciplined failing-test-first flow, on request |
typescript-best-practices |
TS standards applied while reading/editing .ts/.tsx |
Understanding & explanation — know what you're touching
| Skill | What it does |
|---|---|
how |
How does X work — walkthroughs, placement, ownership |
why |
Design rationale, regressions, postmortems |
architect |
Sketch types/signatures/structure before code |
figure-it-out |
Design an auditable playbook for work with no narrower fit |
recall |
Reconstruct working context from session history (via devin_session_* tools) |
reflect |
Mine the transcript for durable learnings; route each to a concrete edit |
teach |
Explain a body of work so a person actually understands it |
bro |
Restate the last message in plain human language |
in-your-own-words |
Check alignment: restate the goal before doing the work |
ketchup / ketchup-video |
Catch you up on what changed since your last message |
technical-writing |
Diátaxis + Google-style + STE layered writing standard |
sharpen |
Make a task deeper, not bigger; rewrite the prompt |
Engineering principles — 25 named judgment calls
attack-the-premise, boundary-discipline, build-the-lever, encode-lessons-in-structure, exhaust-the-design-space, experience-first, explain-the-number, fix-root-causes, foundational-thinking, guard-the-context-window, laziness-protocol, make-operations-idempotent, migrate-callers-then-delete-legacy-apis, minimize-reader-load, model-the-domain, never-block-on-the-human, outcome-oriented-execution, prove-it-works, redesign-from-first-principles, separate-before-serializing-shared-state, sequence-verifiable-units, subtract-before-you-add, test-behavior-not-implementation, type-system-discipline.
Each is a skill that fires at the moment its judgment applies — e.g. laziness-protocol when you're tempted to add abstraction, explain-the-number before trusting any measurement.
Agents — briefs for embedding in child-session prompts: agents/poteto-agent.md, agents/comment-sicko.md.
Automations — automations/benny/: a Slack bug-triage + reproduce-and-fix pack ported to devin_automation_manage, with its own 3 skills (setup-benny, triage-issue-reports, reproduce-and-fix-issues). Copy the pack into a target repo to use it.
Docs — start here:
| Doc | Contents |
|---|---|
docs/guide/01-setup.md … 10-recipes-and-pitfalls.md |
Ten walkthroughs: setup → poteto-mode → understand → design → build → verify → overnight runs → principles → customization → recipes |
docs/harness/devin.md |
Full Cursor→Devin adaptation table + where Devin exceeds Cursor |
docs/harness/other-harnesses.md |
Notes for running these ideas on other agents |
- ✅ All 64
SKILL.mdfiles pass frontmatter validation; zero broken internal links. - ✅ Every Cursor-specific mechanism (Task sub-agents,
/loop,/automate, cloud agent, Custom Mode,.cursor/paths, transcript globbing,origin pr) has been replaced with a real Devin mechanism — audited file-by-file. ⚠️ End-to-end install (skill discovery afterinstall the plugin) is verified to file-structure validity; one real install pass on your machine is the last check — report anything missing as an issue.⚠️ Thebennyautomation pack is documented and ported but hasn't run a live Slack trigger yet.
plugin.json # Devin plugin manifest (agent-plugins.org schema)
.cursor-plugin/ # Upstream pstack manifest, kept for provenance
skills/ # 61 core skills (poteto-mode, swarm, principles, ...)
poteto-mode/playbooks/ # shipping, babysit, autopilot, fleet, ...
agents/ # poteto-agent, comment-sicko
automations/benny/ # Slack triage + reproduce-and-fix automation pack
docs/guide/ # 10-part walkthrough
docs/harness/ # devin.md adaptation table
assets/logo.png # otter eating a sweet potato
Is this official? No. Community port, MIT licensed like upstream. Cognition and Cursor are not affiliated.
Will it track upstream? pstack is Lauren's living skill set — upstream changes don't flow automatically. File an issue (or PR) with the upstream diff you want pulled in.
Does it work in Devin on the web/CLI or only Desktop? The plugin is a directory Devin loads; it works anywhere Devin sessions run, though the workflow assumes the full Devin tool surface (git tools, session tools, review).
Issues and PRs welcome — especially: upstream drift ports, more Devin-native playbooks, and verification reports on platforms you use.
- Lauren Tan (@poteto) — pstack, the skills, the style, and the sweet-potato-eating creature concept that the logo riffs on. Upstream repo.
- Devin port: @Till3005.
MIT — same as upstream. Copyright (c) 2026 Lauren Tan.
