workflow_skill

Overview

This skill designs graph-based workflows for AI agents, including chaining, branching, merging, suspend/resume, streaming, and observability. It guides you from understanding the end-to-end process to producing a clear workflow graph, Mermaid diagram, step table, suspend/resume plan, streaming strategy, and tracing design. The output emphasizes inspectable steps, single-responsibility actions, and durable state for long waits.

How this skill works

I ask focused questions to capture the full process, dependencies, parallelism, human checkpoints, and streaming needs. I map the process to four primitives (chaining, branching, merging, conditions), produce a Mermaid workflow graph and a step-by-step table, and identify suspend/resume points with persistence strategy. Finally I design streaming events and observability traces (OpenTelemetry) and show how workflows compose as tools or steps.

When to use it

• Break a complex agent task into explicit, inspectable steps
• Add structured branching, parallelism, or conditional routing
• Introduce human-in-the-loop suspend/resume checkpoints
• Provide real-time progress (streaming) to users
• Add tracing and observability for debugging and cost analysis

Best practices

• Make each step do one thing — avoid multiple LLM calls per step
• Name steps clearly; names appear in traces and UX
• Persist minimal JSON-serializable state for suspend/resume; don’t keep processes running
• Stream step starts, token output, tool calls, and progress to users
• Instrument every run and step with OpenTelemetry spans and attributes

Example use cases

• Customer support agent: route intent → fetch KB / escalate → generate and validate response
• Order processing: validate order → check inventory (parallel) → charge → notify
• Research assistant: parallel analyses (summarize, extract entities, score) then merge into report
• CI/CD pipeline: code lint → run tests (parallel) → merge decision → deploy workflow
• Payment reconciliation: ingest transactions → match rules → escalate mismatches to human