Designing a Local Agentic Storytelling Pipeline

Overview

This tutorial demonstrates how to build a fully local, API-free agentic storytelling system using Griptape and a lightweight Hugging Face model. We'll create an agent with tool-use capabilities, generate a fictional world, design characters, and orchestrate a multi-stage workflow that produces a coherent short story. By dividing the implementation into modular snippets, we can clearly understand each component as it comes together into an end-to-end creative pipeline.

Environment Setup

We start by setting up our environment:

!pip install -q "griptape[drivers-prompt-huggingface-pipeline]" "transformers" "accelerate" "sentencepiece"

We initialize a local Hugging Face driver and a helper function to display outputs cleanly, allowing us to follow each step of the workflow.

Creating an Agent

Next, we create an agent equipped with a calculator tool and test it:

math_agent = Agent(
   prompt_driver=local_driver,
   tools=[CalculatorTool()],
)
math_response = math_agent.run(
   "Compute (37*19)/7 and explain the steps briefly."
)
show("Agent + CalculatorTool", math_response.output.value)

This validates that our local driver and tool integration work correctly.

World and Character Generation

We construct the world-generation task and dynamically create character-generation tasks:

world_task = PromptTask(
   input="Create a vivid fictional world using these cues: {{ args[0] }}.
Describe geography, culture, and conflicts in 3–5 paragraphs.",
   id="world",
   prompt_driver=local_driver,
)

By defining a reusable function for character tasks, we see the workflow take shape through hierarchical dependencies.

Introducing Stylistic Rules

We integrate stylistic rules and create the final storytelling task:

style_ruleset = Ruleset(
   name="StoryStyle",
   rules=[
       Rule("Write in a cinematic, emotionally engaging style."),
       Rule("Avoid explicit gore or graphic violence."),
       Rule("Keep the story between 400 and 700 words."),
   ],
)

This brings all tasks into a coherent narrative workflow and allows us to run it effectively.

Output and Summarization

Finally, we gather all generated outputs and compute metrics for the story:

metrics = summarize_story(story_text)
show("Story Metrics", metrics)

In summary, we orchestrate complex reasoning, tool interactions, and creative generation using local models within the Griptape framework. This presents a pathway for advanced experiments in local agent pipelines and automated writing systems.

For the complete code, check out the FULL CODES here.