XNews

pstack Author Reveals Engineering Principles Design for Cursor Plugin

pstack Author Reveals Engineering Principles Design for Cursor Plugin

The author of pstack disclosed the design approach behind the Cursor built-in plugin's engineering principles, stating that agent instructions are compressing from lengthy operational steps into short named principles as model capabilities improve. Each principle spans roughly 10–20 lines and can be invoked as a skill via /principle-<name>, while the /poteto-mode mechanism requires agents to name each principle that influenced a decision and the specific choice it changed. The principles derive from the author's own in-session corrections of agents.

Event version 1.

Key insights

Principles as minimal skills

Each pstack engineering principle is about 10–20 lines, a distilled and named good practice rather than a process script. The author expects most skills in pstack to eventually compress into such short principles or workflows, allowing a batch to be deleted.

/poteto-mode decision-naming mechanism

Before multi-step tasks, the agent reads a short principle index. During execution it must name each principle that influenced a decision and the specific choice it changed, generating a "Principles Named or Used" section at the end of the reply.

Principles from real corrections

Most principles come from scenarios where the author personally corrected agents in sessions this year, described as structured sedimentation of lessons.

Principles can be invoked by name

Each principle can be called as a skill via /principle-<name>, and more commonly used by name to steer work. A phrase like "subtract before you add" redirects an agent to delete outdated adapters before designing the rest.

Source posts