Task I/O Schema Design
Define each workflow task's input/output shape explicitly, and separately keep a clear-enough idea of how the task should be accomplished.
In a workflow build process, one task's output becomes the next task's input, so each task's interface must be well-defined before implementation: what information does the task need, what does it produce, and is that I/O structured (with an explicit schema) or free text?
Example (a Shopify plug-in email-generation task): the input schema names
typed fields such as plug-in name, concept, rationale, and a store_id
used to retrieve store details; the output schema is just subject_line and
body, both text. Writing the schema down this concretely, before
implementation, is what lets the task be
implemented and verified in isolation.
Beyond the I/O interface, you also need a reasonably clear idea of how the task should be accomplished — not just "generate content" but, for example, "a fun presentation of the concept designed to appeal to the owner based on values and themes demonstrated on the store's website," which implies pulling content from the web page plus conditioning prompts around it. This how need not be as rigidly fixed as the I/O schema — content is easier to change than an interface — but it should be well-enough defined that you're confident the task is reasonable as scoped; otherwise you end up back at the drawing board rearranging tasks or redesigning interfaces once implementation reveals the task was underspecified. Prefer a non-LLM implementation when traditional software suffices; see non-LLM task implementation.
When a first-draft task is not thoughtful enough, escalate: tighten the prompt, add forced reasoning before tools or chain-of-thought, then Reflexion with format checks, unit tests, or LLM-as-judge analysis before retry. For open-ended tasks, an Assistant–UserProxy pair can supply the conversational partner an expert agent needs. Evaluate each task in isolation with per-task offline harness tests before debugging the whole graph.
Sources
- Prompt Engineering for LLMs (John Berryman), ch. 9