Direct answer
In AI-assisted content, never vary verified facts, source meaning, core definitions, proof boundaries, legal or ethical caveats, audience fit, exclusions or approved claims unless a human explicitly re-approves them. Vary format, length, sequence and examples only inside those limits.
Why this matters
This matters because AI rewriting can quietly change fixed facts, soften caveats or turn approved claims into stronger promises. Naming what must not vary keeps useful transformation from becoming uncontrolled drift.
How to use it
- List the facts and definitions before generation.
- List caveats, exclusions and proof limits.
- Mark transformable elements separately.
- Give the AI permission only for those transformables.
- Reject outputs that change the fixed material.
Example
An AI can turn a long article into a checklist. It cannot turn “reduces hallucination risk” into “prevents hallucination” unless that stronger claim has been proved and approved.
Caveats and limits
Sometimes fixed material must change because the source changed. That should happen through a source update, not casual generation.