design-task

A good CrewAI task description states what to do, how to do it, why it matters, and what constraints apply. Include specific steps, scope limits, and available inputs so the agent produces focused output instead of guessing.

In sequential process, each task automatically receives all prior task outputs. For non-linear dependencies, set context=[prior_task] on the dependent task to explicitly control which outputs it receives.

output_pydantic returns a validated Pydantic model accessed via result.pydantic, while output_json returns a parsed dict accessed via result.json_dict. Both use a Pydantic model to define the schema, and expected_output remains a human-readable string.

Guardrails validate task output before it passes downstream, returning (bool, result_or_error). On failure the agent retries up to guardrail_max_retries. You can chain function-based checks with LLM-based string guardrails.

Literal placeholders appear when a variable in the YAML description is missing from the inputs dict passed to crew.kickoff(). CrewAI uses single braces {variable}, not Jinja2 double braces, and unused inputs are silently ignored.

Do not use human_input to model normal conversational follow-up; that belongs in a conversational Flow calling handle_turn() per message. Reserve human_input for approving or correcting a specific task output before it moves downstream.