Config-Driven Claude: Schema-Based Prompting

When prompts become data structures.

The Problem

Complex AI workflows hardcode logic in prompts:

prompt = """You are a form assistant. If the user is a corporation,
ask about shareholders. If LLC, ask about members. If sole prop,
skip to Schedule C. Also remember to..."""

This doesn’t scale. It’s hard to test. It mixes logic with language.

The Pattern

Separate what Claude says from what Claude does:

{
  "schema_version": "1.0",
  "steps": [
    {
      "id": "entity_type",
      "type": "choice",
      "prompt": "What type of entity?",
      "options": ["corporation", "llc", "sole_prop"],
      "next": {
        "corporation": "shareholders",
        "llc": "members",
        "sole_prop": "schedule_c"
      }
    }
  ]
}

Benefits

1. Testable

def test_corp_path():
    copilot = FormCopilot(config)
    copilot.answer("entity_type", "corporation")
    assert copilot.current_step == "shareholders"

2. Versionable

configs/
  v1.0.json
  v1.1.json  # Added new question
  v2.0.json  # Breaking change

3. Swappable Same engine, different forms:

tax_form = FormCopilot("configs/tax.json")
permit_app = FormCopilot("configs/permit.json")

4. Auditable

copilot.get_path()
# ["entity_type:corporation", "shareholders:3", "board:yes", ...]

Schema Validation

from jsonschema import validate

with open("config_schema.json") as f:
    schema = json.load(f)

validate(instance=my_config, schema=schema)

Catch errors before runtime. No more “prompt didn’t handle this case.”

---

Logic in JSON. Language in prompts. Both versioned.

Source: form-copilot