Understand the system, compare choices, and check the result
pstack
Engineering Work You Can Inspect
Lauren Tan’s engineering workflow plugin combines code investigation, competing designs, parallel implementation and evidence-based review into work with a clear scope, deliverable and way to check it.

Entry points used here
- Cursor: /add-plugin pstack → /setup-pstack → new chat
- Claude Code: installed commands use the /pstack: prefix
- Repository stars
- 14,668
- Version
- 0.15.36
- Checked
- Categories
- Engineering
Start by understanding one concrete entry point
After configuring models, open a new chat and use /how on a real function or user action. Check the file references, call sequence and unanswered questions before handing a change to /poteto-mode. This first task does not require an application test harness.
Pick a task and get started
Install in a Cursor chat opened on the target repository, run /setup-pstack, confirm models and roles, then start a new chat. Select the /how example below and supply a real entry point. Check the explanation and its gaps; this does not establish that the code is defect-free.
Follow data from a real entry point and give the next engineer an explanation they can check.
Install pstack in Cursor
This is a Cursor chat command, not a terminal command. After installing, run /setup-pstack and select a skill in a new session. The source manifest is version 0.15.36; this command does not pin the research commit, so check the version actually installed.
Read installation and entry points/add-plugin pstackConfigure roles, then start a new chat
Run /setup-pstack. It detects models, asks for a reasoning budget and shows a role table before writing the complete user-level ~/.cursor/rules/pstack-models.mdc file. The rule applies to new sessions. Existing settings are read before the file is rewritten; reruns can retain custom roles, while deleting a role line restores its current default. The final offers to create application and formal-verification skills are optional. Skip them for a first read-only investigation. In a new chat, select how from the / menu before entering the task, and confirm that the skill is attached.
Read setup steps and configuration scopeGive your AI this first task
These five editorial examples need your real paths, conditions and commands. Set the permitted files, tools and external actions first. The listed deliverables are expected outcomes, not results tested on this site.
/how I will provide the entry file or symbol for cancelling a task in this repository. Read through one cancellation request: the user action or API entry, task lookup, cancellation signal, worker and final state. If the repository has no matching feature, say so without inventing a path. Assess the scope first. Use one explainer for a narrow question; use the documented 2–4 exploration angles and synthesis only when this genuinely crosses subsystems. Identify real files and functions at each step, with inputs, outputs, state ownership and branches. Add a simple diagram when several participants interact. Distinguish sending a cancellation signal, stopping the worker and persisting the final state. A query inside a loop is a performance lead, not proof of an N+1 problem. Separate confirmed mechanics, inference and paths you could not trace. Return only the explanation, a short file map and open questions. Do not edit files, run the app, access external business systems, commit or open a PR. Before finishing, spot-check the key calls and confirm that the working tree has no new changes.
Check what you receive
- A file- and function-backed execution path explaining data and state ownership.
- A useful file map, a diagram where needed, and explicit gaps; code reading is not presented as runtime verification.
More about this skill
pstack uses /poteto-mode to route engineering work: identify the request, select a playbook, put its steps on a task list, and call focused skills as needed. The current primary skills directory has 58 skills, including 24 engineering principles, while the main workflow has 23 playbooks. You can start with one bounded task whose result you know how to inspect.
The entry points produce different things. /how reads code and explains execution; /architect compares interfaces and module shapes; the bug and performance playbooks require evidence from running software; /interrogate asks reviewers to reproduce failures and has the lead agent rerun them. Understanding code, choosing a design, fixing behavior and proving a finding require different completion checks.
The approach favors less unnecessary code, explicit data ownership and boundaries, and evidence a reviewer can revisit. It is also opinionated: design normally proceeds into implementation, code playbooks proceed toward commits and a PR, and some steps call another plugin. State the writable scope and stopping point before starting a task.
Cursor is the native entry point. Claude Code installs a generated copy with adapter instructions, with differences in command names, model parameters and cloud features. This page starts with a read-only Cursor investigation, then covers design, repair, performance and diff review. All prompts are editorial examples based on the source; none was executed here.
Sources and verification
Sources checked
The pinned README, setup and adapter instructions, main entry, selected skills and references, and generator were read in full. No plugin was installed, third-party code or subagent workflow executed, or target project verified. The source record distinguishes full reading from unreviewed parts of the collection.