Meta-Improvement Context Switch Skill

Purpose: Enable recursive self-improvement by recognising when project work reveals patterns that should become permanent infrastructure, then safely switching context to implement enhancements.

Activation: Detection of generalizable pattern, user says "make this permanent", repeated behaviour across sessions, or self-discovered optimisation.

Token Budget: ~1000 tokens for switch (500 save + 500 restore)

Core Principle

Project work is the laboratory. ~/.claude is the knowledge base.

Every session generates insights:

• Behaviours that work well (should be skills)
• Repeated actions (should be commands)
• Autonomous decisions (should be memories)
• Complex workflows (should be agents)

This skill transforms ephemeral learnings into permanent capabilities.

Trigger Conditions

Activate when:

• Claude makes autonomous decision that should be permanent (like updating next_up.md)
• User says "make this permanent", "add this to memory", "create a skill for this"
• Same pattern detected 3+ times across sessions
• User explicitly praises a behaviour ("good, you learnt this yourself!")
• Efficiency audit reveals repeatable optimisation

The Meta-Improvement Cycle

PROJECT WORK
    ↓
[Pattern Detected]
    ↓
PAUSE: Save project state (session-state-persistence)
    ↓
SWITCH: Navigate to ~/.claude context
    ↓
ENHANCE: Create/update skill/memory/command/agent
    ↓
DOCUMENT: Record WHY the pattern was detected
    ↓
RESTORE: Return to project with todo state intact
    ↓
RESUME PROJECT WORK

Protocol

Phase 1: Pattern Recognition

Signals that indicate generalizable pattern:

• "I decided to do X without instruction because..." (autonomous decision)
• Same action taken in multiple projects
• User positive feedback on unprompted behaviour
• Efficiency gain measurable in tokens or time

Document the WHY:

## Pattern Detection Log
- **What**: [The behaviour/decision]
- **Why**: [Reasoning that led to it]
- **Evidence**: [User feedback, repeated occurrence, efficiency gain]
- **Generalizability**: [How this applies beyond current project]

Phase 2: Context Save

Before switching:

1. Invoke session-state-persistence skill
2. Save current TodoWrite state
3. Note exact resumption point
4. Record any in-progress work

Phase 3: Enhancement Implementation

Navigate to ~/.claude and create appropriate artefact:

Phase 4: Context Restore

After enhancement:

1. Return to project working directory
2. Re-read project state file (next_up.md)
3. Restore TodoWrite state
4. Continue from noted resumption point

Integration Points

With Self-Improvement Engine

• This skill IS part of the 10-step recursive cycle
• Feeds detected patterns to skill generation module
• Triggers documentation updates

With Session State Persistence

• DEPENDS on state persistence for safe context switch
• Calls /save-session-state before every switch

With History Mining Agent

• Can search past sessions for pattern evidence
• Validates "3+ occurrences" threshold

Anti-Patterns

• Do NOT switch context for trivial patterns (one-off decisions)
• Do NOT lose project context during enhancement
• Do NOT create duplicate skills (check existing first)
• Do NOT over-engineer simple patterns
• Do require clear WHY documentation for every enhancement

Quality Gates

Before creating new infrastructure:

1. Threshold: Pattern occurred 3+ times OR user explicitly requested
2. Uniqueness: No existing skill/memory covers this
3. Generalizability: Applies beyond current project
4. Value: Saves tokens OR prevents errors OR improves quality

Example: The next_up.md Pattern

Detection:

• Claude updated next_up.md without instruction
• User said: "good, you learnt this yourself!"

WHY Analysis:

• Previous session established next_up.md as checkpoint file
• Completing Phase 6 was significant milestone
• Efficiency principle: preserve progress to prevent re-work
• Generalizable: ALL projects benefit from state persistence

Enhancement Created:

• Memory: session-state-persistence rule
• Skill: ~/.claude/skills/session-state-persistence/SKILL.md
• Command: /save-session-state

Meta-Enhancement:

• This very skill (meta-improvement-switch) was created because the ACT of creating the session-state-persistence skill was itself a generalizable pattern worth codifying.

Token Savings

Per Enhancement Cycle:

• Avoids re-learning pattern: ~5k-20k tokens (depends on complexity)
• Prevents repeated user reminders: ~500 tokens per occurrence
• Reduces investigation time: ~1k tokens

Long-term:

• Each permanent enhancement saves tokens in ALL future sessions
• Compound effect across projects

Success Metrics

• Enhancement completes without losing project context
• New skill/memory immediately usable
• User confirms pattern correctly codified
• Future sessions benefit from enhancement

Changelog

v1.0 (2025-11-26) - Initial implementation

• Pattern recognition triggers
• Context save/restore protocol
• Integration with session-state-persistence
• WHY documentation requirement

Skill Status: Active Maintainer: Auto-generated from meta-pattern detection Last Updated: 2025-11-26