# claude-kb-workflow

> Open-source knowledge pipeline for Claude Code, for people who maintain many unrelated projects alone. Claude Code's native memory is per-project and per-machine; this toolkit adds the cross-project layer: sessions → LLM wiki → stable KB, with a promotion exam (used by 2+ projects, or a gotcha already paid for, or a declared standard, and always verified in production). Ships orchestrated subagents, production patterns, methodological skills, and a documented restore dry-run that failed. English docs, native Italian included. MIT license.

## Docs

- [README](README.md): overview, quick start, FAQ
- [KB-driven workflow](docs/kb-driven-workflow.md): the three-tier pipeline, the 5 commands, the promotion exam
- [Orchestration](docs/orchestration.md): lead + subagents, three model tiers (Haiku, Sonnet, Opus), mandatory reviewer gate, the corrupted-frontmatter story
- [Knowledge disaster recovery](docs/knowledge-disaster-recovery.md): making a dev setup 100% restorable, the restore dry-run
- [2026 comparison](docs/comparison-2026.md): vs Superpowers/GSD/gstack (complementary, not competing)
- [Patterns](patterns/patterns.md): 16 generalized production patterns
- [Single source of truth](docs/single-source-of-truth.md): which kind of information lives where, and what the other places may hold
- [Incremental review log](docs/incremental-review-log.md): review only the diff since the last logged review

## Italiano

- [Workflow KB-driven](docs/workflow-kb-driven.md)
- [Orchestrazione](docs/orchestrazione.md)
- [Disaster recovery della conoscenza](docs/disaster-recovery-conoscenza.md)
- [Fonte unica](docs/fonte-unica.md)
- [Review log incrementale](docs/review-log-incrementale.md)