nw-forge
nw-forge is a structured five-phase workflow for creating specialized AI agents that progresses through analysis, design, creation, validation, and refinement stages. Use it when building new agents or validating existing agent specifications to ensure focused responsibility, adherence to design patterns, compliance with a 14-point checklist, and extraction of domain knowledge into separate Skills files when content exceeds 50 lines.
git clone --depth 1 https://github.com/nWave-ai/nWave /tmp/nw-forge && cp -r /tmp/nw-forge/nWave/skills/nw-forge ~/.claude/skills/nw-forgeSKILL.md
# NW-FORGE: Create Agent (V2)
**Wave**: CROSS_WAVE
**Agent**: Zeus (nw-agent-builder)
## Overview
Create a new agent using research-validated v2 approach: focused core (200-400 lines) with Skills for domain knowledge.
1. **ANALYZE** — Identify single clear responsibility, check overlap with existing agents, classify type, determine minimum tools needed. Gate: responsibility defined, no overlap, classification chosen.
2. **DESIGN** — Select design pattern, define role and divergent principles, plan Skills extraction, draft frontmatter. Gate: pattern selected, principles drafted, frontmatter ready.
3. **CREATE** — Write agent `.md` using template. Workflow must be numbered task list. Create Skill files if domain knowledge exceeds 50 lines. Gate: agent file written, line count under 400.
4. **VALIDATE** — Run 14-point validation checklist. Check for anti-patterns. Verify workflow is numbered task list, not prose. Gate: all 14 items pass, zero anti-patterns.
5. **REFINE** — Address validation failures. Add instructions only for observed failure modes. Re-measure and re-validate. Gate: all items pass, line count reported.
## Agent Invocation
@nw-agent-builder
Execute \*forge to create {agent-name} agent.
**Configuration:**
- agent_type: specialist | reviewer | orchestrator
- design_pattern: react | reflection | router | planning | sequential | parallel | hierarchical
## Success Criteria
- [ ] Agent definition under 400 lines (`wc -l`)
- [ ] Official YAML frontmatter format (name, description, tools, maxTurns)
- [ ] 11-point validation checklist passes
- [ ] Only divergent behaviors specified (no Claude defaults)
- [ ] 3-5 canonical examples included
- [ ] Domain knowledge extracted to Skills if >50 lines
- [ ] No aggressive language (no CRITICAL/MANDATORY/ABSOLUTE)
- [ ] Safety via platform features (frontmatter/hooks), not prose
## Next Wave
**Handoff To**: Agent installation and deployment
**Deliverables**: Agent specification file + Skill files (if any)
## Expected Outputs
```
~/.claude/agents/nw/nw-{agent-name}.md
~/.claude/skills/nw-{skill-name}/SKILL.md*.md (if Skills needed)
```Review dimensions for validating agent quality - template compliance, safety, testing, and priority validation
Review dimensions for validating agent quality - template compliance, safety, testing, and priority validation
Review dimensions for acceptance test quality - happy path bias, GWT compliance, business language purity, coverage completeness, walking skeleton user-centricity, priority validation, observable behavior assertions, traceability coverage, and walking skeleton boundary proof
Detailed 5-phase workflow for creating agents - from requirements analysis through validation and iterative refinement
5-layer testing approach for agent validation including adversarial testing, security validation, and prompt injection resistance
Architectural style selection decision matrices, trade-off analysis, structural enforcement rules, and combination patterns. Load when choosing or evaluating architecture styles.
Comprehensive architecture patterns, methodologies, quality frameworks, and evaluation methods for solution architects. Load when designing system architecture or selecting patterns.
Canonical AT completeness gate — research-anchored 7-category taxonomy (C1-C7) + 15-item mechanical checklist. Paradigm-neutral. Drives acceptance-designer reviewer verdict deterministically.