English | 한국어
Second Claude Code — Work OS for Knowledge Work
You type one prompt. Researchers crawl 20+ sources. An analyst finds patterns. A writer drafts 3,000 words — and before you even see it, five reviewers are already tearing the draft apart. One checks the logic. Another attacks the weak points. A third fact-checks every number.
One prompt. Full cycle. No duct tape between plugins.
This isn't a coding assistant. It's a work OS — it runs the full knowledge-work cycle autonomously: Plan → Do → Check → Act. Research, analysis, writing, and quality assurance in a single automated loop.
Image created with SceneSteller
Docs · 한국어 문서 · User Manual · 사용 매뉴얼 · Skill Guides · GitHub Issues · 한국어 README
What's New in v1.1.0
- Artifact Viewer — PDCA pipeline outputs now render as an interactive local web UI. Run a pipeline, open
localhost in your browser, and click through phases to see markdown, radar/bar/pie charts, flow diagrams, and syntax-highlighted code — all updating in real-time via WebSocket
- Viewer skill —
/second-claude-code:viewer starts the viewer for any PDCA session. The server auto-stops after 30 minutes of inactivity or when Claude Code exits
- Responsive layout — desktop (768px+) shows a split panel (timeline left, artifact right); mobile (<768px) uses a draggable bottom sheet
What was new in v1.0.0
- 323 tests, green locally — current release verification is
322 passing, 1 skipped, 0 failing
- PDCA Cycle Memory — every cycle now auto-saves to
.data/cycles/, building a persistent learning archive. Start a run, transition phases, end the run — and the system remembers what worked, what failed, and what to watch for next time
- Read-Before-Act — before entering any PDCA phase, the system reads prior cycle insights for the current domain. No cold starts. Every cycle benefits from the last
- Self-Evolution — insights decay over time (stale lessons fade), and gotcha proposals bubble up from recurring failure patterns. The system gets smarter with use
- Domain-aware PDCA —
pdca_start_run accepts a domain parameter (code, content, analysis, pipeline) to enforce specialized stage contracts from the first phase
- 3 new MCP tools —
pdca_get_cycle_history, pdca_save_insight, pdca_get_insights bring the total to 24 tools
- Guardrails on every skill — all 13 skills ship with Iron Laws and Red Flags, plus an anti-fabrication layer in
hooks/lib/fact-checker.mjs for numeric-claim verification
- Stronger gates, fewer false approvals — stage contracts in
config/stage-contracts.json, corrected consensus rounding (2/3 means 2, not 3), score + vote dual gating, and preset-specific thresholds govern phase exits
- Richer cycle outcomes —
pdca_transition can now PROCEED, REFINE, or PIVOT, with max-count caps to prevent infinite loops
- Visual feedback built in — session end emits an ANSI summary box in the terminal and auto-generates dark-theme HTML cycle reports with Mermaid and Chart.js in
.data/reports/
What was new in v0.9.0
- 311-test release baseline — suite sat at 311 total (
310 passing, 1 skipped)
- Domain-aware PDCA starts —
pdca_start_run gained the domain parameter
- Guardrails on every skill — Iron Laws and Red Flags across all 13 skills
- Stronger gates — stage contracts, corrected consensus rounding, dual gating
- Richer cycle outcomes —
PROCEED, REFINE, or PIVOT with bounded retries
- Visual feedback — ANSI summary box and HTML cycle reports
- Reliability upgrades — File Mutation Queue, loop budget caps, iterative compaction, MAD confidence scoring
- Better integrations — optional
mmbridge MCP registration, MMBridge Adapter Protocol, MetaClaw PRM tracking
Quick Start
1. Install
claude plugin add github:unclejobs-ai/second-claude-code
2. Verify — start a new session and look for this in the context injection:
# Second Claude Code — Knowledge Work OS
13 commands for all knowledge work:
Nothing? Run claude plugin list to check.
3. Just talk
Research the current state of AI agent frameworks and write a report
unclejobs-ai