Evaluating AI Skills: Matt Pocock vs. Pstack
Theo - t3.gggo watch the original →
the gist
Treat AI 'skills' (system prompts/markdown instructions) as reference material for building your own workflows rather than plug-and-play templates, focusing on high-impact triggers like 'unslop' to improve model output quality.
The Philosophy of AI Skills
Instead of treating AI skill repositories as rigid sets of instructions to copy-paste, they should be viewed as design patterns for agentic workflows. Blindly installing these files is counterproductive; the real value lies in auditing them to understand how they trigger specific model behaviors. A 'skill' is essentially a markdown file that acts as a trigger—a thumbnail for your agent—to pull in specific context or constraints when the model identifies a relevant task.
The 'Unslop' Effect
One of the most impactful findings is the 'unslop' skill, which forces models to abandon generic, AI-typical phrasing (e.g., 'pivotal moment', 'testament to', 'I hope this helps'). By providing the model with explicit instructions to remove puffery, use direct language, and favor human-like rhythm, the semantic density and readability of AI-generated output improve significantly. This shift transforms the model from a verbose chatbot into a concise, actionable assistant.
Testing and Implementation
Testing these skills does not require full installation. You can simply copy the markdown content into a prompt to see how a model reacts. The speaker emphasizes the distinction between 'user-invoked' skills (manual tools for specific tasks like grilling a project plan) and 'model-invoked' skills (background steering that helps with debugging or architecture). Using a 'grilling' skill—which forces the model to act as a relentless interviewer—is particularly effective for stress-testing project goals and identifying stale documentation or conflicting logic.
Strategic Workflow Integration
Effective engineering with AI involves creating a 'control center' for agents. By using skills like 'Wizard' (to guide humans through tasks agents cannot perform) and 'Triage' (to manage issue state machines), developers can move beyond simple chat interfaces. The goal is to create a coherent workspace where agents manage terminals, git diffs, and checkpoints, allowing for parallel, persistent work across different providers.