[BUG] Skill tool fails to activate plugin-dev skills - missing build.sh script

What’s Wrong?

The Skill tool is executing example code embedded in plugin-dev skill documentation (SKILL.md). Inline examples that demonstrate the command execution syntax are being treated as live directives, causing cascading failures (e.g., missing scripts/build.sh, then missing scripts/analyze.js, etc.). This makes plugin-dev skills unusable.

• I can activate skills from other marketplace plugins without issue.
• The failures are specific to plugin-dev skills, at minimum plugin-dev:command-development.

Steps to Reproduce

Prereqs

• Claude Code with marketplace claude-code-plugins installed
• Plugin location: ~/.claude/plugins/marketplaces/claude-code-plugins/plugins/plugin-dev

Repro

1. Activate the command-development skill:

• Skill --skill plugin-dev:command-development

1. Observe that the Skill tool attempts to run example code from the documentation, first trying a missing build script, then a missing Node script:

• First error (build script missing):
• bash: ~/.claude/plugins/marketplaces/claude-code-plugins/plugins/plugin-dev/scripts/build.sh: No such file or directory
• After creating a dummy scripts/script.sh to bypass one example, a new error appears:
• Node script missing (executed from a different inline example):
• See logs below.

1. Verify the skill content exists and is static markdown:

• cat ~/.claude/plugins/marketplaces/claude-code-plugins/plugins/plugin-dev/skills/command-development/SKILL.md | head -20
• Content renders; there’s no need to “build” anything.

Actual Behavior

• The Skill tool executes documentation examples that use the command execution dialect (! + backticked shell), even when they are clearly presented as fenced code examples in the skill docs.
• This leads to failing subprocess calls whenever referenced example files do not exist (e.g., scripts/build.sh, scripts/analyze.js).

Example error showing execution of an inline example:

EVALUATE: plugin-dev:command-development

  Evaluation Question: Is the audit target a slash command or command specification?

  My Evaluation: YES - The audit analyzes 34 commands in the linko-metaprompt plugin. Request 113 specifically requires validating command frontmatter (name, description, version, allowed-tools), which are command-specific quality criteria provided by this skill.

  Activating NOW:

  ⎿  Error: Bash command failed for pattern "!`node /home/chnzca/.claude/plugins/marketplaces/claude-code-plugins/plugins/plugin-dev/scripts/analyze.js $1`":
      [stderr]
      node:internal/modules/cjs/loader:1215
        throw err;
        ^
      Error: Cannot find module '/home/chnzca/.claude/plugins/marketplaces/claude-code-plugins/plugins/plugin-dev/scripts/analyze.js'
          at Module._resolveFilename (node:internal/modules/cjs/loader:1212:15)
          at Module._load (node:internal/modules/cjs/loader:1043:27)
          at Function.executeUserEntryPoint [as runMain] (node:internal/modules/run_main:164:12)
          at node:internal/main/run_main_module:28:49

This happens despite the skill being static documentation. It also occurs with other example lines, e.g. !bash ${CLAUDE_PLUGIN_ROOT}/scripts/script.sh.

Expected Behavior

• Skill docs should render as static guidance. Example code in SKILL.md must not be executed.
• No subprocesses should be spawned unless a skill explicitly declares a dynamic mode and explicitly opts in to command execution.

Evidence in the Skill File (Examples Being Executed)

Examples from plugin-dev:command-development skill that should be treated as literal examples but are being executed:

```541:549:/home/chnzca/.claude/plugins/marketplaces/claude-code-plugins/plugins/plugin-dev/skills/command-development/SKILL.md
---
description: Analyze using plugin script
allowed-tools: Bash(node:*)
---

Run analysis: node ${CLAUDE_PLUGIN_ROOT}/scripts/analyze.js $1

Review results and report findings.

```555:557:/home/chnzca/.claude/plugins/marketplaces/claude-code-plugins/plugins/plugin-dev/skills/command-development/SKILL.md
# Execute plugin script
!`bash ${CLAUDE_PLUGIN_ROOT}/scripts/script.sh`

There are many similar fenced examples throughout the doc (e.g., ! + backticked commands for git, npm, gh, node, etc.).

Analysis / Likely Root Cause

• The Skill tool appears to apply the “command execution” dialect (the ! backtick runner and possibly inline backticked commands) to skill documentation. It should not parse or execute those inside skills, particularly inside fenced code blocks meant purely as examples.
• The behavior suggests the same interpreter used for runnable slash commands is being reused for skill rendering without a safe “docs-only” mode.

Scope

• Affects at least plugin-dev:command-development. Given similar content patterns, likely affects all plugin-dev skills with executable examples:
• plugin-dev:agent-development
• plugin-dev:hook-development
• plugin-dev:skill-development
• plugin-dev:mcp-integration
• plugin-dev:plugin-structure
• plugin-dev:plugin-settings

Other marketplace plugins’ skills activate normally.

Severity

• Medium-High: Blocks users from accessing the canonical plugin development guidance that other workflows depend on (command audits, agent creation, plugin QA).

Suggested Fix

Engine-side changes for Skill tool:

• Do not execute any ! backtick commands or inline backticked shell within skills by default.
• Treat fenced code blocks as literal text; never execute contents.
• If “live examples” are ever desired, require an explicit, opt-in frontmatter on the skill (e.g., dynamic: true + explicit allowed-tools), with the default being safe, non-executing render.
• Consider filtering the execution dialect entirely for skills and only enabling it in actual slash commands.

Optional docs-side hardening (belt-and-suspenders):

• The plugin-dev skills could add a frontmatter flag (e.g., disable-execution: true) that the Skill tool respects to force safe rendering. But the core fix should be on the Skill tool to avoid executing examples by default.

Environment

• Platform: Linux (WSL2)
• OS: Linux 6.6.87.2-microsoft-standard-WSL2
• Marketplace: claude-code-plugins
• Plugin: plugin-dev
• Skills affected: see Scope
• Claude Code Version: [current version]
• Regression: Yes (previously worked)
• Last known good: 2.0.59
• Current: 2.0.50 (as reported)

Workarounds Tried

• Creating a dummy scripts/script.sh suppresses one failing example, but execution then cascades to other example commands (e.g., missing scripts/analyze.js). This is not a viable workaround.

What Should Happen?

• Skills should render their markdown as static guidance. No example code should be executed when a user opens a skill.

Update

• I replaced all runnable example markers in plugin-dev:command-development from leading ! to the literal text “exclamation_mark” (e.g., exclamation_mark\bash ...\`) in skills/command-development/SKILL.md`.
• After this change, the skill activates successfully and no subprocesses are attempted. This confirms the Skill tool was executing inline examples from the documentation.

Implications

• The execution parser is being applied to skill documentation. Fenced example blocks and inline backticked examples are not treated as inert text in skills.
• Other marketplace plugins’ skills still activate normally; this appears specific to plugin-dev skills that include runnable-style examples.

Recommendation

• Engine: Disable command execution inside skills by default; treat fenced code as literal. If “live examples” are desired, require an explicit opt-in (e.g., dynamic: true + allowed-tools) at the skill level.
• Plugin content: Until the engine fix ships, replace leading ! in examples with a safe literal (e.g., “exclamation_mark”) or another non-executable token across all plugin-dev skills.
• Optional guardrails: Add a CI/lint check in the marketplace to flag !-backticked command patterns in skills/**/SKILL.md.

Current Status

• Workaround verified: plugin-dev:command-development now loads and displays content without errors after replacing ! markers.
• Request: Please fix the Skill tool’s default behavior and/or update plugin-dev skill docs accordingly; similar changes may be needed in the other six plugin-dev skills.

Also encountering this when trying to use command-development

<img width="1000" src="https://github.com/user-attachments/assets/926eec51-8df7-4dd9-8d25-6d6a6a30f3be" />

Replicated on macOS (Darwin 24.6.0) with Claude Code 2.0.60.

Root cause confirmed: The skill files contain ! backtick patterns that are meant as documentation examples showing users how to write their own commands, but the skill loader executes every ! backtick pattern it finds - including ones inside code blocks and example sections.

The problematic patterns in plugins/plugin-dev/skills/command-development/SKILL.md and its referenced files:

Line 547: Run analysis: !`node ${CLAUDE_PLUGIN_ROOT}/scripts/analyze.js $1`
Line 638: Build: !`bash ${CLAUDE_PLUGIN_ROOT}/scripts/build.sh`
Line 639: Test: !`bash ${CLAUDE_PLUGIN_ROOT}/scripts/test.sh`
Line 640: Package: !`bash ${CLAUDE_PLUGIN_ROOT}/scripts/package.sh`
Line 799: - Script: !`test -x ${CLAUDE_PLUGIN_ROOT}/bin/analyze && echo "✓" || echo "✗"`
Line 814: Execute build: !`bash ${CLAUDE_PLUGIN_ROOT}/scripts/build.sh 2>&1 || echo "BUILD_FAILED"`

Plus additional references in examples/plugin-commands.md and references/plugin-features-reference.md.

These are example snippets showing users the syntax for embedding bash in their own commands - they were never intended to be executed. The scripts they reference (analyze.js, build.sh, test.sh, package.sh) don't exist and were never created.

Fix options:

1. Skill loader should not execute ! patterns inside fenced code blocks or example sections
2. Escape the ! in documentation examples so they're rendered as literal text
3. Remove the executable syntax from examples entirely and just show the pattern conceptually

cc @dhollman hope you don't mind being tagged in on this issue 🙇

This issue has been inactive for 30 days. If the issue is still occurring, please comment to let us know. Otherwise, this issue will be automatically closed in 30 days for housekeeping purposes.

Still valid.

This issue has been automatically closed due to 60 days of inactivity. If you're still experiencing this issue, please open a new issue with updated information.

This issue was closed incorrectly despite recent human comments. This behavior of the bot is reported at https://github.com/anthropics/claude-code/issues/16497. Please upvote that issue, so maybe it gets noticed.

This issue has been automatically locked since it was closed and has not had any activity for 7 days. If you're experiencing a similar issue, please file a new issue and reference this one if it's relevant.