Architecture refinement

Architecture Refinement Skill

Meta-skill for maintaining and evolving the AI assistant's cognitive architecture through deliberate documentation and pattern extraction.

Purpose

This skill enables the AI assistant to:

• Recognize when a learning session produces architecture-worthy insights
• Document patterns in the appropriate .github/ location (skills, instructions, prompts)
• Update migration trackers and status tables
• Consolidate knowledge files following KISS/DRY principles

When to Apply

Trigger this skill when:

• A session resolves a recurring problem (document the pattern)
• A skill file is migrated or consolidated (update trackers)
• A new workflow emerges (capture in appropriate .instructions.md)
• User feedback reveals a principle ("don't over-simplify")

Core Patterns

Pattern Recognition Checklist

Before session ends, ask:

Question	If Yes → Action
Did we solve a problem that could recur?	Document in relevant skill file
Did we learn something about the AI assistant's architecture?	Update relevant .github/ documentation
Did a file get created/deleted/consolidated?	Update migration trackers
Did user correct AI behavior?	Add to skill's Anti-Patterns + document principle
Did the skill itself get improved during use?	Commit the refinement immediately

Documentation Location Guide

What You Learned	Where to Document	Audience
Technical skill pattern	.github/skills/{name}/SKILL.md	the AI assistant (AI)
Important concepts for user	.github/skills/{name}/SKILL.md	Human + AI
Process improvement	.github/instructions/*.instructions.md	the AI assistant (AI)
Complex workflow	.github/prompts/*.prompt.md	the AI assistant (AI)
Domain expertise	.github/skills/{name}/SKILL.md	the AI assistant (AI)

Key Distinction:

• .github/skills/ = Operational reference during work (the AI assistant + user)

Consolidation Decision Tree

• >50% overlap with existing skill → Consolidate INTO it, update tracker as "Consolidated"
• <50% overlap → Create new skill, add to tracker as "Migrated"

Pattern Extraction Template

When documenting a learned pattern:

### [Pattern Name]

**Pattern**: [One-sentence description of what to do]

**Example**: [Concrete instance from the session]

**Why**:

- [Reason 1]
- [Reason 2]

**Anti-pattern**: [What NOT to do, if applicable]

Quality Checks

Before committing documentation updates:

• Markdown lint-clean (blank lines around lists, proper headings)
• Tables have header separators
• No orphaned references to deleted files
• Migration trackers reflect current state
• Commit message follows conventional format

Anti-Patterns

Don't	Do Instead
Document every minor fix	Only architecture-worthy insights
Create new file for each learning	Consolidate into existing structure
Wait until end of session	Document as patterns emerge
Over-document obvious things	Focus on non-obvious learnings
Skip human feedback	Capture corrective principles explicitly

User Coaching Learning Loop

User corrections = high-value learning.

Protocol: Acknowledge → Fix → Extract principle → Document in skill → Commit

AI Tendency	User Correction	Extracted Principle
Over-centralize	"Don't dump in one file"	Distribute to appropriate locations
Over-simplify	"You lost context"	Preserve nuance when consolidating
Skip validation	"Did you check it?"	Always verify (lint, count chars)
Assume completion	"What about X?"	Follow through on all aspects
Add diagrams to skills	"KISS them goodbye"	Skills are for AI, not visual learners

---

Connection to Bootstrap Learning

Learn → Coach → Extract → Document → Consolidate