CrewAI

CrewAI is field-based, not message-based. An Agent is described by role, goal, and backstory strings; a Task by description and expected_output. There is no message array and no system/user split to reconcile — you just supply final strings.

So the bridge is direct assignment: render each prompt with Prompting Press and pass the rendered string into the matching field. The rendered output is the render(...).text field — not Prompt.body, which is the raw, un-substituted template.

Render once — don't double-fill

CrewAI runs its own {placeholder} interpolation over these fields when you call crew.kickoff(inputs=...). If Prompting Press already substituted your variables, those placeholders no longer exist, so passing the same variables again to kickoff is at best redundant and at worst confusing. Render with Prompting Press, hand CrewAI final strings, and don’t re-pass those variables to kickoff.

Assign rendered strings to CrewAI fields

integrations_crewai_fields.py

"""Use Prompting Press prompts with CrewAI (Python).

CrewAI is field-based: an ``Agent`` takes ``role`` / ``goal`` / ``backstory``
strings and a ``Task`` takes ``description`` / ``expected_output`` strings.
There is no message array. So the bridge is direct assignment: render each
prompt with Prompting Press and pass the RENDERED string (the ``render(...).text``
field — NOT ``Prompt.body``, which is the raw un-rendered template) into the
matching field.

Note: do NOT ALSO pass the same variables to ``crew.kickoff(inputs=...)``.
CrewAI runs its own ``{placeholder}`` interpolation over these strings; if
Prompting Press already substituted them, the ``{placeholder}`` text no longer
exists and re-interpolation is at best redundant and at worst confusing. Render
once, with Prompting Press, and hand CrewAI final strings. Standalone.
"""

from crewai import Agent, Task
from prompting_press import Prompt
from pydantic import BaseModel


class RoleVars(BaseModel):
    domain: str


class GoalVars(BaseModel):
    topic: str


# Prompt definitions, each rendered to a final string with typed vars.
role_prompt = Prompt(
    {
        "name": "agent-role",
        "role": "system",
        "body": "Senior {{ domain }} researcher",
        "variables": {"domain": {"type": "string", "trusted": True}},
    }
)
goal_prompt = Prompt(
    {
        "name": "task-goal",
        "role": "user",
        "body": "Write a concise brief on {{ topic }}.",
        "variables": {"topic": {"type": "string", "trusted": False}},
    }
)

# Render -> use the .text field (the rendered output string).
rendered_role = role_prompt.render(RoleVars, data={"domain": "robotics"}).text
rendered_goal = goal_prompt.render(GoalVars, data={"topic": "actuator safety"}).text

# Assign rendered strings directly to CrewAI's fields. `llm=` is set so the
# Agent constructs offline without needing a live provider probe; we never call
# crew.kickoff(), so no network / no API key is used.
agent = Agent(
    role=rendered_role,
    goal=rendered_goal,
    backstory="You have published widely on industrial robotics.",
    llm="gpt-4o",
)
task = Task(
    description=rendered_goal,
    expected_output="A one-paragraph brief.",
    agent=agent,
)

print(agent.role)  # "Senior robotics researcher"
print(task.description)  # "Write a concise brief on actuator safety."

# --- assertions (this file is executed by CI) ---

# The rendered .text strings are assigned verbatim to CrewAI's fields.
assert agent.role == "Senior robotics researcher"
assert agent.goal == "Write a concise brief on actuator safety."
assert task.description == "Write a concise brief on actuator safety."
assert task.expected_output == "A one-paragraph brief."

# Guard against the .body-vs-.text trap: .body is the RAW template, .text is
# rendered. A recipe that used .body would ship un-substituted {{ }} text.
assert role_prompt.body == "Senior {{ domain }} researcher"
assert rendered_role == "Senior robotics researcher"
assert "{{" not in agent.role

Each render(...).text result is assigned straight to a field. The example also asserts the .body-vs-.text distinction, since reaching for .body by mistake would ship the raw {{ }} template into your agent.