Your prompt is feeding the model its own cliché

Description and use

The subtitle pipeline produced hooks with the same sentence skeleton, making the output machine‑like. By removing the adjective from the prompt, providing a list of angles, and feeding recent hooks as an avoid‑list, the hooks became genuinely varied.

Problem

An AI‑generated video subtitle pipeline repeatedly produced the same sentence structure, making the output machine‑like.

Outcome

The pipeline now produces hooks with varied sentence structures, eliminating the machine‑like pattern and improving viewer engagement.

Workflow

• 1. Replace the adjective "counterintuitive" with a list of angles for the hook. 2. Add a rule that the hook must be different each day and avoid ending with a question mark unless it is a real question. 3. Implement a function to read the last N hook sentences from the existing log and format them as an avoid‑list. 4. Append the avoid‑list to the prompt before sending it to the model. 5. Ensure the log is read from the same file the pipeline writes to, keyed by date. 6. Return an empty string on any failure to avoid breaking the unattended job. 7. Test by running the old and new prompts on the same input and comparing the hook sentences. 8. Use a dry‑run flag that writes to the log with a [dry] tag so the avoid‑list parser ignores it. 9. Verify the dry‑run output is not fed back into the model as a recent hook.

Evidence

I put the cliché in the prompt and it did what I asked.

Source

Three details that are easy to get wrong: Read from a log you are already writing.

Source