Customization

The WorkflowCustomizer in src/autoflow/catalog/customizer.py lets you fork any catalog template, modify its LLM nodes, evaluate the result, and save it as a custom workflow.

Customization Workflow

Fork a template

Create a copy of any catalog template with a new slug:

python -m autoflow --fork classification --fork-as my_classifier

The fork is saved to workflows/custom_workflows/my_classifier.json with provenance metadata tracking the original template.

Modify LLM nodes

Adjust system prompts, prompt templates, temperature, model, or max tokens on any LLM node:

# Change temperature
python -m autoflow --modify-slug my_classifier --modify-node intake --set-temperature 0.0

# Change model
python -m autoflow --modify-slug my_classifier --modify-node intake --set-model gpt-4o

# Change system prompt
python -m autoflow --modify-slug my_classifier --modify-node intake \
  --set-system-prompt "You are an expert email classifier for IT support tickets."

# Change prompt template
python -m autoflow --modify-slug my_classifier --modify-node intake \
  --set-prompt-template "Classify this ticket into one of: bug, feature, question.\n\nTicket: {{input.message}}"

Evaluate your changes

Run validation and quality scoring on the modified workflow:

python -m autoflow --evaluate-custom my_classifier

This runs:

JSON Schema validation against the workflow schema
Custom structural rules (unique IDs, valid edges, no orphans, variable resolution)
Quality scoring across all 5 dimensions

Managing Custom Workflows

List all custom workflows

python -m autoflow --list-custom

Custom workflow storage

Custom workflows live in workflows/custom_workflows/ and are automatically discovered alongside catalog templates. They include provenance metadata:

{
  "nodes": [...],
  "edges": [...],
  "metadata": {
    "name": "My Classifier",
    "forked_from": "classification",
    "fork_date": "2025-01-15T10:30:00Z",
    "modifications": [
      {
        "node_id": "intake",
        "field": "temperature",
        "old_value": 0.3,
        "new_value": 0.0,
        "date": "2025-01-15T10:35:00Z"
      }
    ]
  }
}

Programmatic Customization

from autoflow.catalog.customizer import WorkflowCustomizer

customizer = WorkflowCustomizer()

# Fork
customizer.fork("classification", "my_classifier")

# Modify
customizer.modify_node("my_classifier", "intake", temperature=0.0)

# Evaluate
result = customizer.evaluate("my_classifier")
print(result["quality_scores"])
print(result["validation_errors"])

# List custom workflows
customs = customizer.list_custom()

Customization

On this page