[BUG] User-level skills from ~/.claude/skills/ appear in Desktop Code tab autocomplete but don't invoke

Resolved 💬 4 comments Opened Mar 19, 2026 by samkorn Closed Apr 17, 2026

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?

Custom user-level skills in ~/.claude/skills/ appear in the Desktop Code tab's / autocomplete menu with the correct description, but when invoked, the SKILL.md content is not loaded into context. Claude treats the /skillname as plain text in the message rather than invoking the skill.

Works correctly in Claude Code CLI — the same skills invoke properly, load SKILL.md content, and Claude follows the instructions.

Environment:

  • Claude Code v2.1.79
  • Claude Desktop (macOS)
  • Claude Max plan, Opus 4.6 (1M context)

Reproduction Steps

  1. Create a skill at ~/.claude/skills/scope/SKILL.md:
---
name: scope
description: "Read-only exploration + planning mode."
disable-model-invocation: true
---

# Scope Mode

You are in **read-only mode**. Do not write, modify, or delete any files.

## Workflow

### Phase 1: Explore
Investigate thoroughly. Read code, search files, check git history.

### Phase 2: Plan
After research, present a plan.
  1. Open Claude Desktop, switch to Code tab
  2. Type /scope — the skill appears in autocomplete with the correct description
  3. Submit: /scope Write a test hello world file
  4. Claude ignores the skill instructions entirely and writes the file

What Should Happen?

Claude should load the SKILL.md content, enter read-only mode, and present a plan instead of writing files — exactly as it does when the same skill is invoked in Claude Code CLI.

What Actually Happens?

  • The skill appears in the autocomplete menu (proving Desktop discovers it)
  • On submission, Claude does not load the skill instructions
  • Claude treats /scope as decorative text and proceeds normally
  • In contrast, /readonly appeared to "work" but likely by name inference — Claude guessed the meaning from the name rather than loading the SKILL.md rules

Related Issues

  • #28047 describes a similar symptom (skill in autocomplete but content not expanded) but involves context: fork and @ references. This bug is simpler: plain SKILL.md, no fork, no @ refs, no special features.
  • #31597 notes skills work in Desktop but aren't shown in Customize panel — different issue (invocation vs. UI listing).

Additional Context

The skill also contains a ! backtick dynamic shell command (!cat ~/.claude/.permission_mode``), but the invocation failure occurs regardless — Claude doesn't load any of the SKILL.md content, not just the dynamic parts.

View original on GitHub ↗

This issue has 4 comments on GitHub. Read the full discussion on GitHub ↗