The development workflow loop for AI coding agents.
npx skills add johnnichev/nv-dev -g -y3 skills that work together for the core development cycle:
| Skill | Command | What it does |
|---|---|---|
| nv:plan | /nv-plan |
Spec-driven development. INTENT → SPEC → PLAN → IMPLEMENT → VERIFY. Headless mode for automation. |
| nv:test | /nv-test |
TDD with property-based testing, mutation testing, dependency-graph test selection. Separate test/implementation agents prevent tautological tests. |
| nv:debug | /nv-debug |
4-phase systematic debugging (observe → analyze → hypothesize → fix). Loop detection (Repeater/Wanderer/Looper). ~95% first-time fix rate vs ~40% ad-hoc. |
These skills aren't independent — they form a loop:
nv:plan → write spec with acceptance criteria
↓
nv:test → tests verify the acceptance criteria
↓ (something fails)
nv:debug → systematic root-cause analysis
↓ (fix found)
nv:test → regression test prevents recurrence
↓
nv:plan → next feature
| Finding | Source |
|---|---|
| Telling agents HOW to do TDD increases regressions by 64% | TDAD paper (arXiv 2026) |
| Dependency-graph test selection reduces regressions by 70% | TDAD paper |
| AI creates 1.7x as many bugs as humans | Stack Overflow 2026 |
| 56% of agent failures are "iterative refinement failures" | Columbia University |
| ~95% first-time fix rate with systematic debugging vs ~40% ad-hoc | Multiple sources |
npx skills add johnnichev/nv-dev -g -yInstalls all 3 skills globally. Works with Claude Code, Cursor, GitHub Copilot, Windsurf, Aider, Gemini CLI, and 9+ other tools.
Open any project and trigger by topic:
/nv-plan # When starting a new feature
/nv-test # When writing tests
/nv-debug # When something is broken
Or just talk naturally — skills auto-trigger on keywords.
- nv:context — Set up context engineering for your repo (the foundation)
- nv:ops — Guardrails, evaluation, multi-agent orchestration
- nv:design — Professional web design with AI
MIT
Built by johnnichev. For engineers who ship.