This is the first post in a series I am calling What I Measured. The rule for the series is simple. I do not get to assert anything I have not counted, and I have to publish the count even when it makes me look bad. This one makes me look bad.
I have spent a year collecting AI tooling. Skills, subagents, MCP servers. Every week something new shows up, I install it, I tell myself it will pay off. Last night I finally counted how much of it I actually use. The answer is: not much.
How did I count this?
I counted invocations against saved session logs, not memory. My assistant writes every session to a transcript file on disk. On 9 September 2026 there were 339 of them, spanning 12 April 2026 to that morning. I wrote a short script that walks every transcript and records, for each installed skill and each installed subagent, how many distinct sessions it appears in. One session, one tally, so a single long session cannot inflate a favourite. Then I compared that list against everything installed on the machine.
The method has one honest limitation and I want it stated before the numbers. It sees only what is written to those transcripts. A skill I invoked in a session that was never saved, or invoked in a way the log records differently, would read as a zero when it was really a one. So treat every count below as a floor, not a ceiling. That is still enough to answer the question I actually cared about, because the gap is not close.
Installed on 9 September 2026: 260 skills, 70 subagents, 10 configured MCP servers. Of the 260 skills, 110 appear in at least one saved session. 150 appear in none. Of the 110 that fired, 74 fired in exactly one session and never again. So 36 skills out of 260 have been used more than once. That is 14 percent.
Which ones actually earned their install?
Nine skills carry almost the whole roster. Ranked by the number of distinct sessions each appears in, out of 339 sessions measured on 9 September 2026: text humanizing 47, knowledge-graph capture 22, technical SEO 18, research scouting 16, local SEO 15, job scouting 13, context training 12, a writing-style pack 10, and the content pipeline 7. Everything else is in single digits or zero.
One name is missing from that chart on purpose. A skill built for a specific client engagement would have placed inside the top ten, and naming it would identify the client, so it is out of the chart and out of this post. Everything else on the list is my own tooling for my own sites, which is why I can publish it.
What never fired at all?
150 skills have never appeared in a single saved session. Not once in five months. They cover advertising platforms I do not buy on, design systems I do not use, and a long tail of speciality packs I installed the week they were released and then forgot about. They cost nothing to keep in disk terms and quite a lot in every other term.
The subagent roster is healthier but not by as much as I expected. 70 installed, 48 appear at least once, 22 never. Four of them do the real work: a search-and-visibility agent in 69 sessions, an efficiency auditor in 30, a routing agent in 24, and a systems agent in 18. After that it drops into ones and twos.
MCP servers were the ugliest number. Ten are configured. When I count actual tool calls rather than the server simply being available, five have never been called once. Half the servers I load into every session, pay context for on every startup, and have to keep working through version bumps, have produced zero calls. Web access, notes, filesystem and browser control carry all of it.
Why do the losers matter more than the winners?
Unused tooling is not free. Every installed skill and every configured server adds description text to the model’s startup context, adds a maintenance surface when a dependency changes, and adds a routing decision the model has to make before it does any work. A roster where 58 percent of entries have never fired is not a rich toolbox. It is noise that the useful 14 percent has to be found inside of.
I have written this exact mistake down before in a different costume. I once built a whole orchestration layer and then deleted most of it after discovering the branches I was proudest of never executed. I have also written about the tool I bought and never opened again, where the real cost was not the subscription but the false sense that the problem was handled. This is the same failure at a bigger scale. I did not install 260 skills because I needed 260 capabilities. I installed them because installing felt like progress.
The pattern underneath all three is that acquiring a capability and using a capability feel identical at the moment of acquisition, and only a count told me apart. That is the whole reason this series exists, and it is the same discipline I apply in the ten SEO beliefs I measured and got wrong, the sibling post in this series.
What am I doing about it?
I am not deleting 150 skills in one pass, because a zero on a five-month window is not proof a thing is worthless. Some of them are genuinely seasonal. What I am doing instead is cheaper and reversible.
- Anything with zero sessions after five months gets unloaded from startup context, not deleted from disk. It can be pulled back on demand.
- An MCP server with zero calls gets disconnected. Five of ten went immediately.
- Anything installed after this date carries a review date. If it has not fired within four weeks of install, it goes back in the box.
- The count reruns monthly. It is a fifteen-line script, so there is no excuse for guessing.
The thing I would tell anyone building an AI stack for their business: your install list is not your capability list. It is the ceiling of what your capability could be. The floor is whatever your logs say you actually ran, and until you count, you are shopping instead of building. Most of the work in the AI systems I build for founders is deciding what not to include, and this count is why I charge for that judgement rather than for the length of the tool list.
There are no external sources in this post. Every number in it comes from my own machine on 9 September 2026, and there is nothing here to cite that is not a file on my own disk.
Count your roster. Publish the losers. Mine is 150.