Add common-mistakes skill from chat history analysis

[sceiler]

Summary

• Adds a new common-mistakes skill with 25+ concrete wrong/correct rules across 8 categories
• Rules were extracted by analyzing user corrections across 500+ Claude Code chat sessions in 21 project directories (~768 MB of JSONL history)
• Complements the existing agent-best-practices skill (high-level principles) with specific, actionable anti-patterns
• Updates README.md to document the new skill

How it was built

1. Wrote a Python script to scan all ~/.claude/projects/ JSONL chat history files
2. Pattern-matched user messages for corrections (e.g., "no, that's wrong", "revert", "why did you", "I said")
3. Extracted ~145 raw correction instances, filtered to ~50 genuine mistakes (excluding false positives like context continuations, feature requests, skill loading messages)
4. Categorized mistakes into 8 themes and wrote concrete wrong/correct examples

Categories

Priority	Category	# Rules
CRITICAL	File & Project Awareness	4
CRITICAL	Multi-Target Instructions	2
HIGH	Deprecated & Outdated APIs	3
HIGH	Output Formatting	4
HIGH	Debugging Anti-Patterns	5
HIGH	Wrong Optimizations	2
MEDIUM	Content & Data Integrity	5
MEDIUM	Security Scope	2

Most frequent mistakes found

1. Not following all instructions — skipping projects, over/under-scoping changes
2. Using deprecated APIs — e.g., Next.js Image priority prop instead of loading="eager"
3. Inconsistent output formatting — markdown in first half, plain text in second
4. Repeated failed fix attempts — trying the same approach 4+ times without changing strategy
5. Applying client-side optimizations to Server Components — memo(), Suspense where they do nothing

Test plan

• SKILL.md follows the required format (frontmatter + sections)
• SKILL.md is under 500 lines (currently ~250 lines)
• metadata.json is valid JSON with required fields
• README.md updated to list new skill
• Manual review of rule quality and accuracy

🤖 Generated with Claude Code