So I tried Matt Pocock's skills...

Theo38:22
0 comments · 0 votesOpen discussionClose discussion
Sign in to join the discussion

    Video summary

    Theo Browne reviews two collections of AI-agent instructions, comparing Matt Pocock's engineering workflows with Pstack's more focused tools. His starting point is selective adoption: use existing work as a reference, examine the Markdown and choose skills that address recurring problems in your own projects instead of installing an entire collection blindly.

    The first extended example examines Pstack's unslop skill, which directs agents to remove vague phrasing, explain concrete actions and preserve a readable voice. Theo Browne compares narrated outputs with and without those instructions and finds the resulting project descriptions and status reports easier to read. He also explains why a skill's description affects when an agent chooses to load it.

    Matt Pocock's grilling workflow is tested against a project with unclear goals and conflicting documentation. The agent asks about users, milestones, terminology, autonomy and product boundaries, exposing decisions the host had not settled. Theo Browne finds the questions useful even when he dislikes answering them, and distinguishes manually invoked planning skills from instructions that can activate during routine debugging.

    The broader review covers workflows for turning decisions into specifications, tickets, implementation and review, plus writing instructions that other agents can follow. Theo Browne values parts of this structure but does not want every prescribed step. He notes that copying a text-only skill can be enough for a trial, while skills with dependencies or scripts require additional setup.

    Pstack examples include generating and comparing multiple attempts, checking what a change might break elsewhere and keeping a reviewable decision trail. Theo Browne emphasizes verifying crucial claims by running code rather than trusting an agent's own account. He also notes that some instructions are tied to Cursor and that parallel attempts can consume substantial inference, reinforcing the case for a small, deliberately managed collection.

    Original YouTube thumbnailWatch on YouTube