Not Available on This Connection

SCIENCE & TECHNOLOGY · SEPTEMBER 12, 2026

On September 11, an email newsletter from Anthropic told me to run a command I had never used: /skill-doctor. The pitch was specific and genuinely useful — run it "to see what each of your skills costs you in context and how often Claude actually invokes it," so I could turn off the ones that had quietly stopped earning their keep. I have collected a lot of skills over four months of daily use with Claude Code, Anthropic's coding assistant. I ran the command. It failed. I ran it two more times, in two more places, and it failed the same way three times running before I actually read the documentation closely enough to see what it had told me to do differently. On the fourth try, done properly, it stopped refusing the connection — and gave me nothing anyway. Four tries in, I still have not once gotten this feature to tell me what it was built to tell me.

What /skill-doctor Is Supposed to Do

/skill-doctor is a real, documented command, not something the newsletter made up. Per skill, it reports a context cost and a usage count, flags the ones that have never been invoked, and says where to turn them off. In an ordinary terminal session the report opens inside the plugin manager's Stats tab; run headless with -p and it prints as plain text instead. It needs a Claude Code build no older than version 2.1.252, and it depends on the session having fetched Anthropic's feature flags at all — one that skips that step, the documentation says plainly, does not get the feature.^[1]

None of that is what happened to me. What happened to me is documented too — just not in the newsletter that sent me looking for the command in the first place.

Why Claude Code Says "Skill Usage Reports Are Not Available on This Connection"

I typed the command into a session running through the Claude app on my phone, and that sentence came back exactly as written, word for word. That much the documentation had already told me to expect, if I had known to go looking for it before I hit it. What it did not prepare me for was what came immediately after, in the same reply, from the assistant itself: "There's no skill-doctor command installed — nothing by that name in the public, example, or plugin skill directories." It recommended a different, real skill called skill-creator as the nearest thing available, and offered to build me a skill-doctor "if that's something you've used elsewhere."

That second sentence is not a hedge or an approximation. It is simply wrong, and it is wrong about the one thing the reply had just gotten right one line above it. The system had already printed the correct, documented explanation for the failure — a disallowed connection type, not a missing feature. The assistant then wrote straight past its own correct output and asserted a different, false cause, with exactly the confidence it would have earned by actually checking.

Four Tries, and the Fourth Wasn't a Success Either

I did not stop at the phone. I opened a terminal from inside that same phone app and tried it there. Failed. I tried it a third time inside the Claude Code session running as a VS Code extension on my own Mac — which, sitting in an editor with an integrated terminal panel right next to it, felt close enough to "the terminal on the machine where the session is running" that I did not think twice about it. Failed the same way a third time.

All three of those were, in a plain sense, my own fault for not reading the third sentence of the documentation closely enough the first time through: a terminal opened from inside the phone app is still the phone app's connection, and a Claude Code session running as an editor extension is not the same thing as a shell. So I opened a real terminal window on my Mac — no phone, no IDE in the way — updated to the latest build, and ran it a fourth time. This time it stopped refusing the connection. It dropped me into the plugin manager's Stats tab, with a note at the top: "/skill-doctor moved — skill usage and context costs now live in this Stats tab." Under a heading reading "Skills loaded this session," it reported: "(no skills loaded)." That was the whole report. No numbers, no list, nothing to turn off. Four tries in, on the one occasion the command finally went where the documentation says it should go, it had nothing to say once it got there.

I don't believe an empty report is the honest shape of the answer here, for what it's worth — not "I wish it were higher," but that this doesn't read like the feature the newsletter described. The pitch was "how often Claude actually invokes" each skill, which reads like a track record built up over time. What came back was scoped to one session that had, at that point, done nothing but start up and update itself — a session in which "no skills loaded" would be true no matter how many skills I own or how often I use them on every other day of the week. Whether that's a bug or just how the feature is scoped, the practical result is the same one I started with: I still don't know which of my skills are worth keeping, which was the entire reason I went looking for this command.

What This Says About the Tool, Not Just the Bug

I keep coming back to something I said to myself the moment I saw that first phone screen, before I understood any of the rest of it: this is exactly the kind of thing Anthropic would want to know about — a specific, reproducible, timestamped account of a real feature failing on a real account, across real machines, with the exact wrong words it said along the way. Not a complaint that the tool is bad; a small, checkable bug report that a company shipping something this fast genuinely needs a supply of. I don't know whether anyone at Anthropic will ever see this particular entry, but writing it down in a form specific enough to be useful, rather than just grumbling that "AI gets things wrong," is the whole point of the exercise. The most useful sentence in that first reply wasn't the wrong diagnosis — it was the correct one sitting one line above it, ignored by the same response that printed it. A tool built to tell me which of my own tools I'm not using would do well to notice that about itself first.

Where I Could Be Wrong

The version banner on that fourth run said quite a lot on its own: Claude Code had just updated itself, "68 features, 412 bugfixes, and 153 other changes," immediately before the command stopped refusing the connection. That update may be the entire explanation for the first three failures — an older build simply not carrying the feature yet — and if so, the Remote Control message on the phone might be the one genuinely separate cause among the three, rather than three instances of the same problem. I can't fully untangle an out-of-date build from a disallowed connection type from the outside, and I'm not certain the VS Code failure and the phone-terminal failure share a cause at all. And on the empty report: it's possible "no skills loaded" is exactly the correct, honest number for a session that has done nothing yet, and the newsletter's framing — not the Stats tab itself — is what overpromised a track record the feature was never built to keep. Either way, the practical fix is the same one I keep landing on: the report should say what window it's counting, so a reader isn't left, as I still am, unable to tell whether an empty answer means "nothing to report" or "wrong question."

Sources

1. Anthropic. skill-doctor — find and disable unused skills. Claude Code documentation. 2026. https://code.claude.com/docs/en/skills#find-unused-skills
2. Anthropic. "Skill usage reports are not available on this connection." Claude Code error reference. 2026. https://code.claude.com/docs/en/errors#skill-usage-reports-are-not-available-on-this-connection
3. Anthropic. Remote Control. Claude Code documentation. 2026. https://code.claude.com/docs/en/remote-control
4. The Librarian's Notebook. My Collaborator Has No Memory Except the Notes I Keep for It — the companion entry, on a confident answer built from a stale one. 2026. my-collaborator-has-no-memory.html