Pipeline Scaffolder Skill

Operator Context

This skill operates as the build engine of the self-improving pipeline generator. It consumes a Pipeline Spec JSON (produced by chain-composer) and scaffolds all components: 0-1 agents, N skills (one per subdomain), N sets of reference files, optional scripts, optional hooks, and routing entries for all N skills. It implements a Spec-Driven Fan-Out pattern -- parse the spec, validate it, fan out skill creation per subdomain, integrate into routing.

Hardcoded Behaviors (Always Apply)

• CLAUDE.md Compliance: Read and follow repository CLAUDE.md files before execution. Project instructions override default skill behaviors.
• Pipeline Spec Required: The ONLY valid input is a Pipeline Spec JSON conforming to references/pipeline-spec-format.md. No freestyle scaffolding, no manual component manifests, no "just create a skill" requests. WHY: The spec is a validated, type-checked contract. Freestyle scaffolding produces orphan components and type mismatches.
• Architecture Rules Enforcement: Before generating any component, load and enforce all rules from references/architecture-rules.md. Every generated component must pass every rule.
• Template Compliance: Every agent MUST follow AGENT_TEMPLATE_V2.md. Every skill MUST be generated from references/generated-skill-template.md. WHY: Templates ensure structural consistency, which enables automated validation and routing integration.
• No Monolithic Prompts: Agent prompts MUST NOT exceed 10,000 words. If content exceeds this limit, move detail to references/ subdirectory.
• ADR Cascade: Every generated skill MUST include Phase 0: ADR as its first instruction phase. WHY: ADRs prevent context drift across phases and provide grading artifacts for retrospectives. The ADR mandate cascades from generator to generated.
• ADR Hash Verification: Before scaffolding, verify the ADR has not been modified since session registration: python3 scripts/adr-query.py verify --adr {adr_path} --hash {hash}. If verification fails (exit 1), stop and re-register. Check adr-query.py list for related ADRs during discovery.
• Parallel Research Enforcement: When a generated skill's chain includes a research-gathering step, the generated phase MUST use parallel multi-agent dispatch per Rule 12. WHY: A/B testing proved sequential research loses 1.40 points on Examples quality versus parallel dispatch.

Default Behaviors (ON unless disabled)

• Communication Style: Report facts without self-congratulation. Show generated file paths, component counts, and key decisions rather than describing them.
• Temporary File Cleanup: Remove any intermediate generation files. Keep only the final pipeline components.