Claude Usage Patterns Across kulisama81 Repos
Summary of how Claude has been used across 10+ repositories from Dec 2025 to Apr 2026.
Key Takeaway
Claude integration evolved through three distinct phases — from simple co-authoring to ticket-driven development to fully automated multi-agent pipelines — over roughly 4 months. The trajectory shows increasing delegation and trust, with quality maintained through automated gates rather than manual review.
Three Phases of Evolution
Phase 1: Direct Collaboration (Dec 2025 — Jan 2026)
Simple CLAUDE.md with “DO NOT” rules, direct co-authored commits, primarily Sonnet 4.5. SnowCheck was the first project (90% Claude commits). Key innovation: “update CLAUDE.md whenever Claude makes a mistake.”
Phase 2: Ticket-Driven Development (Jan — Feb 2026)
Added tk ticket tracker, GitHub Actions for mobile ticket creation, more detailed CLAUDE.md files with 16+ gotchas rules. Projects like favcollege and anichat-visual-novel-system built rapidly.
Phase 3: Multi-Agent Pipelines (Mar — Apr 2026)
Full orchestration with 6 specialized agents in apibstudyguide. Automated quality gates via hooks. Analytics-driven prioritization. Dropped Co-Authored-By in favor of structured ticket-referenced commits. tkt replaced tk with MCP server integration.
Statistics
- ~83+ explicit Claude co-authored commits across all repos
- 10 repos with some Claude involvement
- 8 repos with CLAUDE.md files
- Highest density: SnowCheck (90%), anichat (100%), Pollitique (80%)
Notable Patterns
- Model selection became intentional: Opus for planning/orchestration, Sonnet for execution — mirrors PM/engineer split
- Single-day builds: anichat and Mimigame were built in one day each
- Self-enforcing quality: PreToolUse hooks block ticket closure without clean UX review
- Living documentation: “DO NOT” lists grew from experience, encoding mistakes as institutional knowledge
See also: claude-integration-timeline, multi-agent-architecture, anichat-visual-novel-system, provider-pattern, sentiment-analysis