Refresh

Refresh Context & Audit Workflow

Skill Metadata

• Name: refresh
• Type: Procedural workflow for context loading and compliance audit
• Auto-trigger: Session startup (light mode), User signals (full mode)
• User signals: "check your work", "that's not right", "read the guidelines", "look again"

Purpose

Reload behavioral guidelines, report available capabilities, and audit recent actions for compliance with meta-CLAUDE.md requirements. Use this when Claude violates guidelines or at session start to establish context.

When to Use

Automatic (Light Mode - Session Startup):

• Re-read meta-CLAUDE.md
• Report available skills and MCPs
• No audit of recent actions

Manual Trigger (Full Mode - User Invoked):

• User types /refresh
• User says "check your work" or similar
• Claude detects it may have violated guidelines
• After corrective feedback from user

Workflow

Phase 1: Reload Guidelines (All Modes)

1. Read ~/dotfiles/ai/docs/meta-CLAUDE.md
   • Report last modified date
   • Confirm loaded successfully
   • Note key sections: Communication, Git Workflow, Code Change Approval

Phase 2: Inventory Capabilities (All Modes)

2. List Available Skills

   • Check <available_skills> in system prompt (what Claude can actually invoke)
   • Scan ~/.claude/skills/*/SKILL.md on filesystem (what exists on disk)
     • Use find -L to follow symlinks (skills directory is symlinked)
     • Command: find -L ~/.claude/skills -name 'SKILL.md'
   • Report both counts to detect registration issues
   • Format: skill-name: brief description

3. Check MCP Servers

   • List configured MCP servers
   • Show tool count per server
   • Note: tree-sitter, repomix, context7, memory, postgres-mcp

Phase 3: Audit Recent Actions (Full Mode Only)

4. Check Last Git Commit (if any in current session)

   • Run git log -1 --format='%s%n%n%b'
   • Verify commit message format matches meta-CLAUDE.md:
     • Brief descriptive title
     • Optional body
     • Ends with "By PB & Claude"
     • NO "Co-authored-by:" lines
     • NO tool attribution (🤖 Generated with...)
   • If violations found: Offer to amend commit

5. Check Recently Modified Files

   • Run git diff --name-only HEAD~1..HEAD (if committed)
   • Or use session memory of files modified
   • For each file, verify header format matches language:
     • Python: # Author: PB and Claude
     • Markdown: <!-- Author: PB and Claude -->
     • Etc.
   • If violations found: Offer to fix headers

6. Check for Unauthorized Changes

   • Review what was changed vs. what was requested
   • Flag if changes went beyond explicit request
   • Remind: "ONLY modify what is explicitly requested"

Phase 4: Report (All Modes)

Light Mode Output (Startup):

Context loaded
├─ meta-CLAUDE.md (2025-11-06)
├─ Skills: 6 registered, 8 on disk
│  ├─ Registered: commit, code-change-approval, code-explore
│  │              context7, memory-augmented-dev, new-file
│  └─ On disk only: postgres-optimization, refresh
├─ MCPs: 5 connected (tree-sitter, repomix, context7, memory, postgres)
└─ Working in: ~/docs

Full Mode Output (Manual/Audit):

Context refreshed - Full audit

Guidelines: meta-CLAUDE.md (2025-11-06) ✓

Skills: 6 registered, 8 on disk
├─ Registered (invokable):
│  ├─ commit: Git commit workflow
│  ├─ code-change-approval: File modification approval
│  ├─ code-explore: Codebase exploration
│  ├─ context7: Library documentation
│  ├─ memory-augmented-dev: Persistent memory
│  └─ new-file: File creation with headers
└─ On disk only (not loaded):
   ├─ postgres-optimization
   └─ refresh

MCPs: 5 connected
├─ tree-sitter (26 tools)
├─ repomix (7 tools)
├─ context7 (2 tools)
├─ memory (14 tools)
└─ postgres (10 tools)

Recent Actions Audit:
├─ Last commit: [commit hash]
│  └─ ✗ Violation: Contains "Co-authored-by:" line
│  └─ Should I amend this commit?
├─ Modified files: src/foo.py
│  └─ ✓ File header present and correct
└─ Change scope: ✓ Only requested changes made

Working in: ~/docs (clean)

Phase 5: Corrective Actions (Full Mode, If Violations Found)

7. If commit message violations detected:

   • Ask: "Should I amend the last commit to fix the message format?"
   • If yes: Use git commit --amend with corrected message
   • Check authorship first (never amend others' commits)

8. If file header violations detected:

   • Ask: "Should I add/fix headers in [file list]?"
   • If yes: Use Edit tool to add proper headers

Communication Style

• No flattery or praise - Just report facts
• Concise - Structured output, not verbose explanations
• Action-oriented - If violations found, offer specific fixes
• Professional peer - "Should I fix this?" not "Let me fix that for you!"

Implementation Notes

Trigger Detection:

• Detect phrases like: "check your work", "that's not right", "read the guidelines", "what are your instructions", "check meta-CLAUDE"
• When detected: Execute full audit mode automatically

Startup Behavior:

• Light mode on session start
• Don't audit actions from previous sessions
• Just establish current context and capabilities

Audit Scope:

• Only audit current session actions
• For commits: Check last 1 commit if made in this session
• For files: Check files modified in this session
• Don't audit historical work

Skill Sources:

• <available_skills> in system prompt = skills Claude can actually invoke via Skill tool
• Filesystem ~/.claude/skills/*/SKILL.md = skill definitions on disk
• IMPORTANT: Use find -L when scanning skills directory (it's a symlink to dotfiles)
• Discrepancies indicate skills that exist but aren't registered (need debugging)

Error Handling

• If meta-CLAUDE.md not found: Report clearly, suggest location
• If skills directory empty: Report and continue
• If MCP servers unavailable: Note which ones, continue
• If git not available or no repo: Skip git-related checks