pstack gates its parallel agents behind a verification skill your project has to generate

The load-bearing command here is /create-verification-skill, which gives a project its own mechanism for proving behavior ^[10]. Tan's working definition of verification is concrete: the agent performs the task, interacts with the actual product, inspects the result, recognises failure, tries again, and keeps going until it can demonstrate success ^[11]. That loop is not a prompt trick; it is a program driving your app and reading the result.

Which is where the invisible constraint sits. If the app cannot be driven and inspected by something other than a human, the skill has nothing to call, and per theneuron.ai's account Tan follows that all the way down, arguing teams should weigh their ability to debug and control an application when choosing a technical stack ^[13]. That sentence is easy to skip past, but it carries the argument. It says the agent workflow is downstream of an architecture decision most teams made years ago.

The adoption number is the one to be careful with. Cursor's engineering team had used Tan's personal skills 10,000 times in a single week at the point she announced open-sourcing them ^[2]. Spread across seven days that is roughly 1,430 invocations a day ^[20]. Team size is not given, so the per-engineer rate is unknown, and a usage count says nothing about defects caught. For the figure to mean anything in your shop you would need an application an agent can exercise end to end, and engineers who read the playbook rather than the diff.

The counts also need a caveat. The repo figures are 23 and 23, while the Cursor marketplace lists 47 skill entries and two specialised subagents ^[5], and the explainer does not say how those sets map onto each other ^[22]. At version 0.15.1 ^[3] that is a snapshot of a pre-1.0 plugin, not a stable inventory.

Then the ordering. Lesson 5 of Tan's guide, as theneuron.ai lays it out, is that parallelism comes after trust ^[15]. The tools that make parallelism easy are already in the box: /swarm splits slices of a problem across agents, /arena runs competing attempts at the same problem, /interrogate sets multiple models on a finished diff ^[14]. All three multiply output. Run them before the verification skill exists and you have bought more diffs against the same human reading capacity, which is the failure the writeup opens with, five agents producing code faster than anyone can read it, while every green check still raises a new question ^[19].

Cost is acknowledged and, in the material supplied, unquantified: the explainer carries a section headed "Yes, this can be absurdly expensive" ^[17] and the text breaks off before the numbers. Multi-model interrogation and fan-out are the obvious line items, and there is no figure here to argue with.

The repo also ships /unslop, which cleans up AI-flavored prose ^[16]. A project that includes a tool for the smell of its own output is at least self-aware.

The license, the file counts, and the router behaviour are documented: /poteto-mode picks the workflow, loads the playbook, builds the work sequence, and invokes skills ^[7]. The claim underneath, that these 23 principles encode senior judgment rather than one engineer's habits, is not something this material lets you check. That test happens in a codebase that is not Cursor's.