Agent skills under .agents/skills are not listed in the <available_skills> list of the system prompt. This makes it hard for the agent to discover them.
My Fix is just to maintain a list of them into a rule. This requires additional tokens to create and maintain
Steps to Reproduce
Prompt the Agent with:
“Whats skills do you have available to you? You are not allow to use a grep or look anything up.”
Expected Behavior
Agent also is able to list the skills from .agents/skills.
Hey, thanks for the report. Before we escalate, we should rule out a couple common causes since this kind of symptom is often config related.
Can you check three things:
File structure. Skills should be placed like .agents/skills/<skill-name>/SKILL.md (it must be SKILL.md inside a subfolder, not just a .md file in the root skills/). Discovery looks for files with that exact name in subdirectories.
Skill frontmatter. Open one SKILL.md and check if it has disable-model-invocation: true. If it does, the skill is intentionally hidden from the model (it shows up in the slash menu and Settings, but it won’t appear in <available_skills>). That was the cause in a similar thread: Why agents can not see my skills in .cursor/skills folder?
Whether the skills show up elsewhere:
Cursor Settings (not VS Code settings) > Rules, Skills, Subagents, are your skills listed there?
In chat, type /, do they show up in the slash menu?
If they show up in Settings and in the slash menu but not in <available_skills>, this looks like an injection bug. If so, please share the contents of one SKILL.md (remove any secrets) and I’ll file a ticket. If they don’t show up in Settings either, it’s a discovery issue, and we should start by checking the folder structure.
Hey thanks for the reply. This is how one skill looks like on my side: SKILL.txt (1.6 KB) - I had to rename it to a .txt extenstion to be able to upload it here.
Thanks, I can see both the file and the Settings screenshot. The structure and frontmatter look good: .agents/skills/<name>/SKILL.md, and there’s no disable-model-invocation.
One last sanity check before I report the bug internally: in the chat, type / and see if your skills show up in the slash menu (they should appear there among the slash commands). If they do, it’s an injection bug (discovery works, but they don’t get included in the <available_skills> part of the system prompt). If they don’t show up in the slash menu either, then it’s a runtime-side discovery issue.
In both cases I’ll report it. I just want to point to the right layer in the ticket.
Small nit about the test itself: asking the agent what it can do isn’t the most reliable signal. The model sometimes doesn’t mention available skills even if they’re in the prompt. But in your case, with 8 skills and a clear “no skills available”, this looks like a real issue, not a hallucination.
Awesome, this is the final confirmation. Skills are being discovered correctly (they show up in Settings and in the slash menu), but they’re not getting injected into the <available_skills> section of the system prompt. So this is a bug in the injection stage, not discovery.
I filed an internal report with a link to this thread and your SKILL.md example. I can’t share an ETA for a fix yet, but we’ll post an update here once we have one.
For now, your workaround with a rule that lists the skills is totally fine if you don’t want to wait.
