Steering coding agents with named principle skills in pstack

In pstack, engineering principles work as a shared vocabulary between the person and the coding agent. pstack is an open-source skill pack for Cursor agents published by lauren, whose README calls them the skills the author uses every day at Cursor (repository). It ships 24 principles, each a short skill holding one rule; in the thread lauren puts them at roughly 10 to 20 lines each. The main /poteto-mode skill carries an index of them, reads it at the start of every multi-step task, applies the ones the task triggers, and is told to name in its reply each principle that shaped a decision and the specific choice it changed. That instruction is why pstack replies often end with a list of the principles named or used. Anyone who wants more direct control can invoke a single rule as a skill, with a /principle- command followed by that principle’s name.

The project's principles guide says you don't invoke principles so much as steer with their names. Because each name points to a complete rule the agent has already read, a short phrase can redirect it more precisely than a paragraph of new instructions. If an agent claims success because the build passed, "apply prove it works" plus a request to run the real flow sends it back to check the actual result rather than a proxy. If it is about to add a fourth adapter next to three old ones, "subtract before you add" tells it to delete the obsolete ones first. The required citation also gives the reader something to check: the guide says a principle cited without a decision behind it is the sign the agent name-dropped the rule instead of applying it.

lauren treats the prose itself as provisional. The thread says many principles came from lauren's own sessions correcting agents this year, that most pstack skills are expected to shrink into short principles or step lists as models improve, and that some can probably be deleted. The author also generally prefers moving guidance into deterministic code, since lint rules, strong types and good API design beat comments, agents.md files, rules and skills. pstack's own meta principle, Encode Lessons in Structure, says the same thing: turn advice you've repeated into a lint, check or script. In an earlier post lauren argued that letting agents verify their own work comes first and that skill packs are an optional extra. On cost, the thread says the main expense comes from large swarms of subagents, and that you can tell /poteto-mode not to spawn them.