The hardest part of getting value from AI coding tools isn't finding good techniques. It's getting a whole team to actually use them. Many AI workflow improvements live and die in one engineer's head — a clever prompt here, a custom instruction there — and never compound across the organization.
A few weeks ago at PlantBid, one of our engineers dropped a link in our AI Tools Discord channel: Matt Pocock's skills repo. Today, nearly every engineer on the team is using these skills daily. That speed of adoption is worth examining, because it says something about how AI practices should spread inside an engineering org.
What skills are, briefly
Skills are markdown files that agents like Claude Code load on demand. Each one encodes a specific way of working — a procedure the agent follows when the situation calls for it. Pocock, who many will know from Total TypeScript, has open-sourced a collection of them that reflect how he actually works with coding agents.
The one that stuck hardest for us is Grill with Docs. Before executing on a task, it pushes the agent to do two things: grill the developer — surface every assumption and ambiguity as a direct question before writing any code — and then record the answers in internal documentation that the skill builds and maintains alongside the codebase.
The grilling alone is worth the install. Anyone who has worked with coding agents knows the failure mode: the AI fills gaps in your instructions with guesses, takes a few weird liberties, and you discover them three files deep in the diff. Grill with Docs front-loads that negotiation. The agent asks; you answer; then it builds. We've seen meaningfully less rework as a result — not because the model got smarter, but because the process stopped letting it guess.
But the documentation half is the deeper win, and it's the part the name undersells. Every decision you confirm gets written down in docs that live with the repo. The next session — yours or a teammate's — consults those docs first, instead of re-asking the same questions or, worse, guessing differently this time. Your answers stop being ephemeral chat history and become a decision record the agent actually reads. Over weeks, the project accumulates context that makes every subsequent agent run sharper, and that survives both context windows and personnel changes.
Why this spread when other tips don't
I've watched plenty of workflow improvements fail to spread over my career. A clever workflow demoed in a meeting impresses everyone and changes nobody's habits.
Skills spread because they're artifacts, not advice:
- They're files. You install them, you don't memorize them. The barrier to adoption is a clone, not a behavior change.
- They're inspectable. An engineer can read exactly what the skill instructs the agent to do, judge it, and modify it. Trust comes from transparency.
- They're versionable. When someone improves one, everyone gets the improvement. The team's AI workflow compounds instead of fragmenting.
This is the same principle behind the AI playbooks we maintain internally: institutional knowledge that lives in files the tools can read beats institutional knowledge that lives in people's heads. Grill with Docs operationalizes that principle at the level of individual decisions; skills as a format extend it from what the team knows to how the team works with agents.
The leadership angle
None of this came from a top-down mandate. It came from an engineer sharing something useful in a channel we'd deliberately created for exactly that purpose. If you lead a team using AI tools, the cheap, high-leverage move is to build the conditions for this kind of spread:
- Give AI discoveries a home. A dedicated channel signals that experimenting with these tools is part of the job, not a side hobby.
- Prefer shareable artifacts over verbal tips. Encourage people to package what works as skills, rules files, or playbooks — things a teammate can adopt in thirty seconds.
- Let adoption be voluntary and earned. Our engineers use these skills because they work, not because I told them to. That's the only kind of adoption that lasts, and usually the only kind that matters when talking about workflow improvements.
Bottom line
Go try Pocock's skills — start with Grill with Docs. But the bigger takeaway isn't any single skill. It's that the teams getting the most from AI tools are the ones treating their working practices as shared, versioned artifacts rather than individual craft.
Worth asking of your own team: when one of your engineers figures out a better way to work with AI, does it spread — or does it stay with them?