Workflow Build Process

Build an LLM workflow in five ordered steps — define the goal, specify tasks, implement tasks, wire the workflow, then optimize — for modularity.

A basic LLM workflow — individual tasks may use an LLM, but routing between tasks is a traditional fixed pipeline, distinct from LLM-driven workflow routing — trades generality for strength on complicated, multi-step work where a conversational agent runs into structural limitations; see generality-strength tradeoff for when that tradeoff is worth making. It is best built in five ordered steps:

  1. Define goal — identify the purpose: what output or change should the workflow accomplish?
  2. Specify tasks — break the workflow into tasks that, executed in order, achieve the goal. For LLM-based tasks, identify needed tools plus each task's inputs and outputs (see task I/O schema design).
  3. Implement tasks — build tasks as specified, with clearly defined I/O, each verified to work correctly in isolation.
  4. Implement workflow — connect tasks into a complete workflow, adjusting individual tasks as needed so they function correctly in the full context.
  5. Optimize workflow — improve quality, performance, and cost.

Workflows earn their appeal through modularity: breaking a complex problem into components makes it easier to build, and when something breaks it's easier to reason about and isolate the failing piece — the same isolation benefit named generally by task decomposition prompting and by AI pipeline orchestration's step-to-step format checks. This modularity is why fixed workflows are the shape to prefer under progressive agent architecture; reach for LLM-driven workflow routing only once that strength genuinely still needs open-ended agency.

Sources

  • Prompt Engineering for LLMs (John Berryman), ch. 9