pstack Playbooks
Every playbook in poteto's pstack, one lane each, plus the entry that picks one and the fallback skill.
Each playbook is one lane, read left to right. Point at a step or a labelled connector and its details open in the panel on the right. Click to pin them; click again to release. Esc closes the panel. The address of the page records a pinned selection, so a link to it shares it. A step with no icon is done from the playbook's own text.
Understand or planBuildCheckCommit, PR, mergeRecordA human decides herenext step / loop backstarts a new passour inference, not stated in the playbooka decision on this connector; point at itprompt files the model reads (skills, playbooks, references)programs that run (scripts, CLIs, built-in commands)
Entry
How a request reaches a playbook
cursor/plugins @2f87c85The human types
/poteto-mode; the model matches the request to one of 22 titles and copies that playbook's steps into the todo tool. PR status and landing are separate requests.How it is chosenthe model reads a list of titles; no code
"Match the task to a playbook and open its steps at playbooks/<title>.md, with the title in kebab case. Copy its steps into the todolist tool first. Large, cross-cutting, or unattended work, or a task no playbook fits, routes to the
/figure-it-out skill."2PlanMatch playbookModel picks one of 22 titles
Skills and Files 1
3PlanFigure it outNo title fits; large or unattended
Skills and Files 1
4BuildRun playbookFollow its steps from the todo list3
Skills and Files 3
5ShipOpening a PRLast step of code-changing playbooks
Skills and Files 1
6ShipBabysitOnly when asked for PR status
Skills and Files 1
7ShipShippingOnly when asked to land
Skills and Files 1
figure-it-out (the fallback)
cursor/plugins @2f87c85/figure-it-out · a skill, not a playbook
"When the task matches no playbook, design one. The deliverable before any code is the workflow itself."
Ends witha reply and a decision trail
Reply: the playbook you designed, the rigor level and why, the decision-trail path, what is verified against the predicate, and what is still open.
1PlanFrameDone predicate, scope, rigor level3
Skills and Files 3
4BuildRun loopHypothesis, smallest change, measure, keep or revert3
Skills and Files 3
6LearnAudit trailA row as each step lands
Skills and Files 1
Tools and Hooks 1
7CheckVerify wholeAgainst the Phase A predicate
Skills and Files 1
Understand
Investigation
cursor/plugins @2f87c85"You own the answer. Plan, route, write." Read-only: a cited explanation or a recommendation, not a code change.
Ends witha reply; no PR
Reply: the investigation output. For "are we sure?" answers, your real judgment with reasons; push back if the premise is wrong. "No PR, no babysit, no
/architect unless the investigation precedes a code change."Runtime forensics
cursor/plugins @2f87c85"You own the diagnosis. Instrument the live process, don't theorize from source." The deliverable is a cited diagnosis, not a fix.
Ends witha diagnosis; no fix unless asked
Reply: the signal captured, the reduced finding, how you proved the mechanism, the source location, artifact paths. "Hand back to Bug fix or Perf once the cause is known."
1PlanCapture signalCPU profile, heap snapshot or CDP trace2
Skills and Files 2
2PlanReduce artifactFind the smoking gun3
Skills and Files 3
Trace forensics
cursor/plugins @2f87c85"You own the diagnosis from the artifact." The capture already exists: "read it, don't re-run it."
Ends witha diagnosis; no fix unless asked
Reply: the artifact and format, the reduced finding, the source location, the artifact paths, and whether a paired capture confirmed it. "Route to Bug fix or Perf issue once the cause is known."
1PlanLoad artifactIdentify the format; right tool3
Skills and Files 3
2PlanMake queryableDump into sqlite, one row each
Tools and Hooks 1
Change code
Bug fix
cursor/plugins @2f87c85"You own this task. Plan, review, verify." "Every shipped line traces to runtime evidence… When evidence refutes a hypothesis, revert what it motivated."
Ends withOpening a PR
Reply: what was broken, root cause, fix, how you verified. "Paste failing-then-passing repro output verbatim."
2PlanFind causeBinary-search hypotheses with runtime evidence2
Skills and Files 2
Tools and Hooks 1
3PlanPlan fixArchitect if it crosses a function
Skills and Files 1
5CheckVerifyOriginal repro passes, same surface
Skills and Files 1
6ShipOrder commitsFailing repro first, fix on top2
Skills and Files 2
Tools and Hooks 1
7ShipOpening a PRRun the shared PR playbook
Skills and Files 1
Feature
cursor/plugins @2f87c85"You own the design. Plan, review, verify." Delegate implementation; stay in the lead.
Ends withOpening a PR
Reply: what you built, what you chose and why, the throughput checkpoint, open decisions. Tables for design alternatives.
1PlanTrace subsystemRead what the change touches
Skills and Files 1
3PlanThroughput checkpointFour todo items; n/a needs a reason
Skills and Files 1
4BuildDelegate buildSubagent, or arena for several shapes5
Skills and Files 5
6ShipOrder commitsSmall ordered commits; stack follow-ups
Skills and Files 1
Tools and Hooks 1
7CheckInterrogateOnly if the design is contested
Skills and Files 1
8ShipOpening a PRRun the shared PR playbook
Skills and Files 1
Refactoring
cursor/plugins @2f87c85"You own the contract. The structure changes. The behavior does not." A redesign routes to Feature; large structural work routes to
/figure-it-out.Ends withOpening a PR
Reply: the structure that changed, the pin you held it against, the equivalence proof, the reader-load delta, what shipped and what got reverted. No new behavior.
1CheckPin behaviourA characterization test before any move
Skills and Files 1
2PlanName structureWhat the code is missing
Skills and Files 1
4BuildSubtract firstDelete dead weight before adding2
Skills and Files 2
6CheckProve unchangedOn the real artifact2
Skills and Files 2
8ShipCommits and PRSubtraction, reshape, cleanup; then PR2
Skills and Files 2
Tools and Hooks 1
Perf issue
cursor/plugins @2f87c85"You own the measurement story." "Tie every fix to a measurement, don't read source instead of measuring." Sustained improvement belongs to Hillclimb.
Ends withOpening a PR
Reply: baseline number, post-fix number, delta, artifact path.
1CheckBaseline traceCapture and vet the number3
Skills and Files 3
3BuildPlan and delegateArchitect if needed; subagent builds4
Skills and Files 4
4CheckPost-fix traceCapture after the change2
Skills and Files 2
5CheckCompare artifactsJSON to sqlite, then diff
Tools and Hooks 1
7ShipOpening a PRRun the shared PR playbook
Skills and Files 1
Hillclimb
cursor/plugins @2f87c85"You own the metric and the experiment's integrity." For sustained improvement of one measurable thing. "One change, one measurement, keep or revert."
Ends withOpening a PR
Reply: the metric and target, baseline to final with the percent delta, iterations run (kept vs reverted), each accepted fix on one line, the decision.tsv path, and the best idea you would try next.
1PlanFix metricWorkload, one metric, stop predicate
Skills and Files 1
2CheckFreeze harnessProve sensitivity; baseline and gate2
Skills and Files 2
3LearnOpen logdecision.tsv, one row per attempt2
Skills and Files 2
Tools and Hooks 1
7CheckKeep or revertPast noise and gate green2
Skills and Files 1
9ShipOpening a PRRun the shared PR playbook
Skills and Files 1
Prototype
cursor/plugins @2f87c85"You own the design decision, not the code. The prototype is a throwaway instrument." Speed over polish; no planning, no tests.
Ends witha decision; the real build follows Feature
Reply: the variants explored, the evidence, tradeoffs, your recommendation, and the scratch path. "Say plainly that the prototype is throwaway."
1PlanScope decisionNo decision means no prototype
Skills and Files 1
4BuildOne switcherVariants side by side, labelled
Skills and Files 1
5CheckObserveScreenshots or logged timing
Skills and Files 1
6PlanRecommendDecision plus the throwaway artifact
Skills and Files 1
Visual parity
cursor/plugins @2f87c85"You own pixel-exact equivalence. The baseline is the spec. You do not touch it." Equivalence is verified by image diff, not by eye.
Ends withOpening a PR per component
Reply: components migrated, the diff result for each, the baseline harness location, what is left.
1CheckBaseline harnessScreenshots of every state, first
Skills and Files 1
4BuildMigrate componentOne at a time, own worktree3
Skills and Files 3
5CheckImage diffA nonzero diff is a fail
Skills and Files 1
Tools and Hooks 1
6ShipOpening a PRRun the shared PR playbook
Skills and Files 1
Skills
Authoring a skill
cursor/plugins @2f87c85"You own the skill's voice." "When in doubt, delete. Keep only prose that changes a decision."
Ends withOpening a PR
Reply: summary of the skill, key design decisions, validation notes.
1BuildWrite skillCursor's built-in create-skill3
Skills and Files 3
4ShipOpening a PRRun the shared PR playbook
Skills and Files 1
"You own the experiment design. Plan, blind, run, synthesize." Candidates must not learn they are being tested.
Ends witha recommendation; no PR
Reply: variant under test, rubric, per-candidate notes, judge's verdict, your synthesis, and a recommendation for whether to promote the variant.
1PlanFrameVariant, success, judge-only rubric
Skills and Files 1
5CheckBlinded judgeDifferent model family, labels only
Skills and Files 1
Pull requests
Opening a PR
cursor/plugins @2f87c85"Invoked at the end of every other playbook." Ends at an open, ready PR. " Opening a PR does not start a babysit."
Ends witha ready PR; no babysit
No Reply line. The PR body is the output: Why, What changed, Scope, Tradeoffs, Blast Radius, Verification.
1ShipWorktreeOff trunk; unrelated work out
Skills and Files 1
Tools and Hooks 1
2ShipSmall commitsRebase into landable commits
Skills and Files 1
Tools and Hooks 1
3CheckDeslopStrip slop from the diff2
Skills and Files 2
5ShipOpen readyNot a draft; stack by base branch
Tools and Hooks 1
Babysit
cursor/plugins @2f87c85"You own the merge frontier. Declare a mode, clear one PR at a time, stop where the human's call begins."
Ends withmerge-ready, never merged
No Reply line. "Owner approval is a wait. Babysitting never authorizes merging."
1PlanDeclare modedrive, threads-only or check
Skills and Files 1
4BuildReview threadsFix real, dismiss noise, escalate
Skills and Files 1
Tools and Hooks 1
5CheckCIA flake earns one fresh build
Tools and Hooks 1
Shipping
cursor/plugins @2f87c85"You own what lands. Verify each PR independently, land only the verified run from the root, then keep your hands off the stack."
Ends withmerged PRs up to the ceiling
Reply: the verified run and its ceiling, each PR's verdict and who produced it, what landed and how you confirmed it, and what the next gap needs.
1CheckVerify each PROne non-author verifier per PR2
Skills and Files 2
Tools and Hooks 1
3CheckRe-check verdictsgit patch-id at verdict versus head
Tools and Hooks 1
4ShipMerge lowestOne PR; read the state back
Tools and Hooks 1
5ShipRetarget nextBase to trunk; force-with-lease2
Tools and Hooks 2
Long runs
Multi-phase plan
cursor/plugins @2f87c85"You own the plan, not the code. The plan is a checklist an owner runs box by box and the operator audits from the evidence." "The plan is the deliverable. Do not implement."
Ends witha plan file; waits for the operator
Reply: the plan path, the PR ids with their dependencies and the review-gated set, and what stays unproven.
1PlanSkip checkOne or two files: no plan
Skills and Files 1
2PlanSettle by prototypeRun it; ask only preference calls
Skills and Files 1
3PlanWrite plan filePR order, dependency graph, checkboxes
Skills and Files 1
Autonomous run
cursor/plugins @2f87c85"You own the exit condition. Define done, then drive to it without stopping."
Ends withthe predicate met
Reply: the exit condition, iterations run, what landed, what was discarded, final predicate state.
1PlanExit conditionA checkable predicate, stated first
Skills and Files 1
2PlanWake mechanism/loop with a watcher or heartbeat
Skills and Files 1
Tools and Hooks 1
3BuildSmallest changeWhat the evidence justifies
Skills and Files 1
5LearnCheckpoint rowWhat changed; did the predicate move
Skills and Files 1
Tools and Hooks 1
Autopilot-full
cursor/plugins @2f87c85"You own the verdicts, never the PRs. One owner carries each PR from build to merge, and nothing merges until your swarm, which never wrote the code, has proven the code that lands."
Ends withowners merge verified PRs
Reply: catch the operator up per the
/ketchup skill: each PR's state and verdict, what merged, countersigns and why, open operator gates, parked follow-ups, with links to each PR and the trails.1PlanOperator gatesThe grant covers only queue items
Skills and Files 1
2BuildSpawn ownersOne cloud agent per PR42
Skills and Files 4
Tools and Hooks 2
3BuildCode-ready headDeslop, no-comments, rebase, report2
Skills and Files 2
Tools and Hooks 1
4CheckSwarm-verifyIndependent verifiers, once, at that SHA3
Skills and Files 3
5ShipMerge prepRebase on trunk; wait for green CI2
Tools and Hooks 2
6ShipOwner mergesOnly what the swarm verified
Skills and Files 1
Tools and Hooks 1
7CheckRoot audit/loop 1h probes every agent
Skills and Files 1
Tools and Hooks 1
Autopilot-stack
cursor/plugins @2f87c85"Run the Autopilot-full playbook, but do not merge."
Ends withverified, unmerged PRs
Reply: list each PR with its verdict, in place of what merged.
2BuildStack if dependentBase on the parent PR's branch
Skills and Files 1
Tools and Hooks 1
3BuildRestackA new lower head moves those above
Tools and Hooks 1
Session
Session pickup
cursor/plugins @2f87c85"You own the resume point. Read the prior trail, don't redo it."
Ends witha routed playbook
Reply: where the prior agent stopped, what you inherited vs redid (ideally nothing redone), the resume point, and the outcome.
1PlanLocate trailTranscript, cloud URL or branch3
Skills and Files 3
2PlanRebuild stateBranch, landed work, todos, decisions
Tools and Hooks 1
Pause safely
cursor/plugins @2f87c85"You own a clean stop. Leave a checkpoint a cold-start agent can resume from." Explicit only: on "keep going" or "don't stop", do not pause.
Ends witha pause, not a final report
Reply: where you are in the loop, what is on disk versus still in your head (paths, no diff dumps), the commits you made and whether the tree is clean, and the first action on resume.
Worktree cleanup
cursor/plugins @2f87c85"This is the one playbook that deletes user state with no code review to catch a slip, so the gates above are the review."
Ends withpruned worktrees; space reclaimed
Reply: df -h / before and after with space reclaimed, the worktrees pruned, and a one-line reason for each held back.
1PlanSnapshot and auditdf -h, then worktree-audit.sh2
Skills and Files 1
Tools and Hooks 2
3CheckVerify usageSubagents read the transcripts2
Skills and Files 2
5ShipPrunegit worktree remove --force
Tools and Hooks 1
My reading
- A playbook is not a skill. It is a plain file of 8 to 25 lines (320 lines for all 22) with no frontmatter, inside the poteto-mode skill folder. The plugin manifest registers only skills/ and agents/, so the harness does not know playbooks exist: no description to match on, no slash command, no hook. The only way in is poteto-mode, which a human must type (disable-model-invocation: true). Its last section tells the model to open one file with an ordinary file read and copy the steps into the todo tool. Subagents never pick one: "The chat that receives /poteto-mode owns the playbook, the todo list, and every Task spawn."
- Choosing a playbook is the model reading one sentence. poteto-mode ends with the 22 titles, comma-separated, under "Match the task to a playbook". There is no trigger table, no keyword list and no code. The fallback, figure-it-out, is also chosen by the model.
- Three programs appear in all 22 playbooks, and none of them decides anything: watch-pr reports a PR's state through exit codes, worktree-audit.sh suggests a bucket and never deletes, log.sh appends a row. Every gate, cap and step order on this page is prompt text.
- A human is needed in few places: typing /poteto-mode, asking for a babysit or a landing, the operator's go in Multi-phase plan and the autopilots, irreversible writes, and uncommitted work in Worktree cleanup. Everything else follows "Just do it".
- The same three rules repeat across the code playbooks: reproduce or measure on the real surface first; "Inconclusive or wrong-surface is not a pass"; a change that did not help is reverted, not left in.
- One contradiction: Opening a PR says it is "Invoked at the end of every other playbook". Seven playbooks say "Run Opening a PR" as their last step, the two autopilots use it inside, and the rest end with a reply, a plan, a diagnosis or a merge.
- Worth taking for hyper: a stop predicate that pairs a target with a floor on attempts (Hillclimb); a verdict that covers a patch, checked with git patch-id, not a commit (Shipping); pin the behaviour before moving structure (Refactoring); "No decision means no prototype"; a plan whose every PR has unit, live and perf boxes with the trunk baseline measured first (Multi-phase plan).