poorrican/claude-code-internals-inspection

Claude Code Internals Inspection

Use this when the user asks to find Claude Code's internal prompts, output style instructions, official plugin prompt recreations, or similar local Claude Code implementation details. Prefer updating the broader claude-code skill when it is writable; this skill exists for prompt/style inspection details.

Workflow

1. Identify the installed Claude Code executable and version:

   command -v claude
   claude --version
   readlink -f "$(command -v claude)"
   

2. Check official plugin recreations first. Claude Code migrated/deprecated some output styles into official plugins implemented as SessionStart hooks:

   ~/.claude/plugins/marketplaces/claude-plugins-official/plugins/*output-style/
   

   The prompt is usually in hooks-handlers/session-start.sh, emitted as JSON under:

   hookSpecificOutput.additionalContext
   

   Parse the heredoc JSON and report the additionalContext, not the shell wrapper.
3. For built-in/native output style strings, inspect the installed binary:

   strings -n 5 /path/to/claude | grep -i -C 20 'Explanatory\|Learning\|outputStyle'
   

4. If exact extraction matters, do not rely only on strings: Claude Code native binaries may contain mixed UTF-8 and UTF-16LE string chunks. Read bytes around the match and decode both encodings. Look for style metadata near names such as Explanatory, Learning, Proactive, descriptions, and headings like # Explanatory Style Active or # Learning Style Active.
5. Report provenance clearly:
   • "official plugin recreation / SessionStart hook" for plugin prompts
   • "native built-in prompt strings from installed Claude Code version X" for binary-extracted strings

Known locations

• Explanatory plugin recreation: ~/.claude/plugins/marketplaces/claude-plugins-official/plugins/explanatory-output-style/hooks-handlers/session-start.sh
• Learning plugin recreation: ~/.claude/plugins/marketplaces/claude-plugins-official/plugins/learning-output-style/hooks-handlers/session-start.sh

Pitfalls

• Do not claim the plugin prompt is identical to the native built-in prompt unless you verified both. The Learning plugin may explicitly say it combines the unshipped Learning style with Explanatory functionality.
• Do not copy README summaries when the user asked for prompts; extract the actual injected context from the hook or binary.
• If a search returns too much noise, narrow to ~/.claude/plugins/marketplaces/claude-plugins-official/plugins and then to *output-style*.