It wasn't the tool. It was one person who refused to go back.

Matt Pocock ships one of the most-installed skill suites in the ecosystem: 9.2 million installs on skills.sh, with four skills in the all-time top ten. This week he asked his own adopters a question vendors usually pay research firms to answer: what was the thing that made your org go "yes, let's do this"?

I pulled all 78 replies and quote-tweets and read them. Bias disclosed up front: these are answers to "why did you say yes," from people who said yes. What follows is how adoption looks when it works, not how often it works. With that on the table: the replies converge hard, and not on features.

Nobody describes a bake-off run by management. The comparisons happen bottom-up, after one person has already decided. One reply compresses the whole dataset into a sentence: "usually it wasn't the tool. it was one person who refused to go back." The long version repeats across companies: "Started using it myself and results talk loud. Ran several workshops with members of each team. In the end everyone was excited. It's now company standard." Another org got there via a demo at the company town hall. Another runs an internal AI guild and settles its arguments with the tooling.

And they adopted one skill, not a suite. Thirty of the 78 replies name the same entry point: /grill-me, a skill that interrogates you about your plan before the agent writes code. The moment shows up in reply after reply ("the plan before vs after grilling felt so much clearer"), and the rest of the suite arrives later, as the obvious next step. The catalog is what you explore after you already believe.

When rivals do get named, six replies cite the same reason for bouncing off them: weight. Too much ceremony, too many tokens, "bibles worth of markdown". And the sleeper reason simplicity wins is in this reply: "We can read each skill and understand what they do (as opposed to the crazy blackboxes out there)." A skill a skeptical senior engineer can read start to finish is a skill that survives review. Readability is the security review.

The most revealing reply came from someone who had already written his own skills, nearly identical to Pocock's, and couldn't get them adopted: "my org is naturally resistant to change… Putting a trusted authority such as yourself as the author lightens people's resistance. Your skills mostly worked better than mine anyway so it was a win win." He concedes Pocock's versions were somewhat better. The resistance dropped when a known name went on the file. Inside a company, provenance is the argument.

So if you're the person trying to get your team onto skills, the thread is your playbook: use it yourself until the results are undeniable, lead with one hero skill instead of a library, pick something readable enough to survive a skeptic, and if your own name isn't enough yet, borrow the authority of someone whose is.

The same thread also shows where the playbook runs out.

The happiest adopters are straining. One: "the increased amount of skills and constant changes is making me a bit overwhelmed. it's hard to catch up." That's update fatigue from following a single author. Another adopted the skills, then deleted most of them: "those skills are good for you but not for us :) so we created our own versions, inspired by yours": a fork, orphaned from every upstream fix the moment it was made. And nobody asked about measurement, but one adopter volunteered the state of the art anyway: "We haven't done any analysis of before after but overall the feel of improvement is valid." Champion to workshop to company standard, and the entire chain runs on vibes and hand-carried files.

Outside the thread, the base rate is sobering. One replier, relaying what friends see at several companies: "monorepos with hundreds of commits per week with high AI usage and 0 skills in repo." Secondhand, one datapoint, hold it loosely. But it matches the replier who demos the hero skill and watches most people "just nod their head and continue vibe coding away."

Skills are spreading through companies the way open source did a generation ago: bottom-up, champion-led, one hero project at a time, on the champion's personal credibility. Open source stayed stuck at that stage until the boring infrastructure showed up around the enthusiasm. Skills are at the stage before that infrastructure: when the champion changes jobs, the workshops stop, the forks drift, and nobody can say what the rollout actually improved, because nobody measured.

The town-hall demos, the workshops, the "here, copy my folder": all of it is one person's unpaid labor compensating for infrastructure that doesn't exist. Reply after reply in that thread is the same person at a different company. Everything their company's agents do well traces back to them, and it walks out the door when they do.