A joke on Threads became a working automation: submit a grammatically challenged song title, and GPT-4o returns a pompous correction from the Grand Curator of the Institute for Musical Grammar Rectification.
The workflow is intentionally small: Form Trigger to OpenAI to Edit Fields to Form Ending. The useful part is the prompt architecture, structured JSON output, and the many tiny lessons hiding inside a silly experiment.
Collects a song title and optional band name.
Runs GPT-4o with a persistent Curator persona.
Parses the response into structured fields.
Displays the decree back to the submitter.
The System message establishes the Curator as a consistent character. The User message sends each new song title as a fresh request. That split keeps the personality stable without repeating the full persona instructions every time.
The model returns raw JSON with four fields: original, corrected, band_correction, and curator_note.
The Respond to Webhook node did not work with the n8n Form Trigger, so the workflow needed a Form Ending node. Fixed vs Expression mode also caused the AI to receive literal template text instead of the submitted song title.
The lesson: test each node individually, especially when a failure looks like a creativity problem but is really a field-mapping problem.
The next iteration would add a second AI call that writes a fake concert-hall program description for the corrected song title. It also needs guardrails before wider use, because a public input field is always an invitation to discover why governance exists.