Skip to main content
Fluid Forge

ChatGPT Builder Checklist

Use ChatGPT's GPT editor in configuration view, not the conversational builder.

1. Create The GPT

  • Open Explore GPTs
  • Select Create
  • Switch to the configuration view

2. Fill The Basic Fields

Suggested name:

FLUID Forge Contract Architect

Suggested description:

Drafts and reviews FLUID 0.7.5 contracts for Forge, asks provider-specific follow-up questions, and requires validation before anything is treated as final.

Instructions:

  • Paste the full contents of GPT_INSTRUCTIONS.md

Conversation starters:

  • Paste the entries from CONVERSATION_STARTERS.md

Recommended model:

  • Pick the strongest reasoning-capable model available in the GPT editor on the day you configure it

3. Set Capabilities

Recommended v1 setup:

  • Data Analysis: On only if users will upload CSVs, JSON schemas, dbt artifacts, or structured metadata
  • Web browsing: Off
  • Image generation: Off
  • Actions: Off
  • External tools/apps: Off

Reasoning:

  • The packet is intentionally knowledge-first
  • Turning off extra tools reduces drift and makes failures easier to trace back to instructions or knowledge files

4. Upload Knowledge

Upload the files listed in UPLOAD_MANIFEST.md.

Guidelines:

  • Upload the curated packet, not the whole repo
  • Keep the knowledge base text-first and compact
  • Remove old or conflicting drafts before publishing updates

5. Sharing

Use one of these:

  • Private to you
  • Workspace/internal only
  • Unlisted internal link

Do not publish to the GPT Store for this version.

6. Preview Smoke Tests

Run these in Preview before sharing:

  1. Generate a FLUID 0.7.5 contract for a local CSV cleanup pipeline from these fields: order_id, customer_id, amount, order_date.
  2. Generate a FLUID 0.7.5 contract for a GCP customer metrics table with GDPR residency and AI restrictions. The GPT should ask which region before it drafts.
  3. Review this contract and tell me what will fail validation: then paste an intentionally broken contract
  4. Upload a small schema-like CSV or JSON artifact and ask the GPT to map it to exposes[].contract.schema

7. Acceptance Rule

The GPT should never treat output as final until someone runs:

fluid validate contract.fluid.yaml --strict
fluid plan contract.fluid.yaml --provider <provider> --out plan.json

Add --check-sovereignty to the plan when the contract has a sovereignty block.

If validation has not been run yet, the GPT should label the result Draft.