What problem does it solve? Building multi-step agent pipelines by hand is error-prone: invalid DAG edges, missing required config, mismatched route conditions, and silent runtime failures. This Skill guides the composition of valid, production-ready workflow definitions for the Resparkable orchestration engine, covering all 19 step types, error strategies, versioning, and validation rules. ## Core Features & Use Cases - DAG Composition: Wire steps across 19 step types — llm_call, route, guard, evaluate, judge_call, human_approval, rag_retrieve, external_call, parallel, supervisor, report, and more — with correct config and conditional edges. - Quality & Audit Patterns: Add post-hoc supervisor audits with evidence-cited verdicts and deterministic Markdown reports, with run-time opt-out toggles for operators. - Validation & Gotcha Avoidance: Apply backend validator rules (required config, reachability, cycle detection), semantic validation of agent/capability/model references, and documented failure modes like LLM-mode guards hallucinating on closed-set checks. - Use Case: Build a customer-support workflow that classifies intent with a route step, retrieves knowledge-base context via rag_retrieve, drafts a reply, gates it through a schema-mode guard, pauses for human approval, then sends a notification with a supervisor verdict attached. ## Quick Start Ask the agent to build a workflow that routes incoming support requests to the right agent, adds a human approval gate before sending, and ends with a supervisor audit and report.
