[BUG] Skill invocation fails when referenced mid-sentence with "use" prefix

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

When a user references a skill with /skill-name mid-sentence (e.g., "use /total-recall I thought we pulled this from slack info before"), Claude Code fails to invoke the skill. Instead of calling the Skill tool, the model generates a conversational response claiming the skill isn't available or "set up," despite the skill being listed in the available skills and fully functional.

The model acknowledged afterward that the skill was available and there was no technical barrier to invocation. The failure appears to be a behavioral issue where mid-sentence skill references are treated as conversational context rather than tool invocation triggers.

What Should Happen?

Any message containing /skill-name (where skill-name matches an available skill) should trigger skill invocation via the Skill tool, regardless of where it appears in the sentence or what words surround it (e.g., "use", context sentences before/after).

The system prompt states: "When a skill matches the user's request, this is a BLOCKING REQUIREMENT: invoke the relevant Skill tool BEFORE generating any other response about the task."

Error Messages/Logs

No error messages. The model simply chose not to invoke the skill and instead responded

"I don't have total-recall set up to search past conversations effectively right now."

This was factually incorrect. The skill was available and functional.

Steps to Reproduce

1. Configure a project with multiple skills available, including one called total-recall

1. In a long conversation with significant context, send a message that references the skill mid-sentence:

``
use /total-recall i thought we pulled this from slack info before should have been written to the roster
``

1. Observe that the model does NOT invoke the skill and instead generates a text response

Expected: The model invokes the total-recall skill via the Skill tool.
Actual: The model skips skill invocation and generates a conversational response.

Note: When explicitly called out and asked to retry, the model successfully invoked the skill on the next attempt using the exact same skill name.

Claude Model

Opus

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

2.1.63

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

iTerm2

Additional Information

• Actual skill and installation performed using downloaded tool here.
• This occurred deep in a long session with significant context (QA lead operations, calendar checks, Linear queries, multiple skill invocations already completed in the session)
• The model had already successfully used other skills earlier in the same session (gcal, linear-integration, bloon)
• The skill total-recall was listed in the system reminder of available skills at the time of the failed invocation
• Context pressure from a long session may be a contributing factor, as the model may be more likely to take shortcuts or make assumptions when context is large
• The user had to explicitly ask "explain exactly what you mean" to surface that the model had simply chosen not to try, at which point it admitted the error and successfully invoked the skill