pre-session-end

Pre-Session-End Skill

Run this skill before ending a Claude Code session to capture learnings and propagate useful settings.

When to Activate

• User says: "done", "wrapping up", "ending session", "goodbye", "that's all", "I'm finished"
• Significant work completed without explicit continuation request
• User explicitly invokes /pre-session-end

Workflow

Step 0: English Correction

Before proceeding with other steps, provide a corrected or more natural version of the user's original prompt from this session.

Format:

## English Correction

**Original:** [user's original prompt]
**Corrected:** [corrected/more natural version]
[If changes are significant, add brief explanation of what was changed and why]

• Only include this section if corrections are needed
• Focus on grammar, word choice, and natural phrasing
• Keep explanations concise

Step 1: Analyze Session for Learnable Patterns

Review the chat history and identify:

Skill Candidates

Patterns that could become reusable skills:

• Repeated workflows (e.g., "always run tests before commit")
• Problem-solving approaches that worked well
• Domain-specific knowledge applied multiple times
• Code patterns or architectures used consistently

Command Candidates

Operations that could become slash commands:

• Multi-step sequences performed repeatedly
• Common parameter combinations for tools
• Project-specific workflows

Agent Candidates

Specialized task patterns that could become agents:

• Complex multi-tool operations
• Tasks requiring specific expertise or context
• Recurring analysis or review patterns

Step 2: Present Extracted Learnings

For each candidate, present to user:

**Potential Skill: [name]**
- Pattern observed: [description]
- Would save: [how it helps]
- Proposed location: ~/.claude/skills/learned/[name].md

Create this skill? [Yes/No/Edit]

Step 3: Propagate Project Permissions to Global

1. Read project settings: .claude/settings.local.json
2. Read global settings: ~/.claude/settings.json
3. Identify permissions in project but not in global

For each permission found only in project settings:

Categorize the permission:

• Generic tool permissions: Bash(git:*), Bash(npm:*), Bash(docker:*) → likely useful globally
• Domain-specific web: WebFetch(domain:docs.example.com) → ask user if generally useful
• Project-specific paths: Contains project directory paths → skip (not portable)
• MCP server permissions: mcp__* → ask user (may be project-specific)

Present each selectively:

**Permission found in project settings:**
`Bash(helm template:*)`

This permission allows: Running helm template commands
Currently: Project-only
Recommendation: [Useful globally / Project-specific]

Propagate to global settings? [Yes/No]

Step 4: Apply Approved Changes

1. For approved skills: Create files in ~/.claude/skills/learned/
2. For approved permissions: Update ~/.claude/settings.json
3. Report what was applied

Step 5: Session Summary

Provide a brief summary:

## Session Wrap-Up Complete

**Learnings captured:**
- [X] Created skill: skill-name
- [X] Created command: command-name
- [ ] Skipped: pattern-name (user declined)

**Permissions propagated:**
- [X] Bash(helm:*) → global
- [ ] WebFetch(domain:internal.corp) → kept project-only

**Ready to end session.**

Important Notes

• Never auto-apply: Always get user confirmation for each change
• Preserve existing: Don't overwrite existing skills/settings without explicit approval
• Be selective: Only propose truly useful patterns, not every repeated action
• Explain value: For each proposal, explain why it's worth saving