1. The refinement loop
Conversational editing is iterative by design. The first result is a draft, and correcting it is usually one sentence — which is faster than undoing a good edit to fix one detail inside it. The conversation carries context, so a follow-up is understood against the edit that already exists rather than as a fresh instruction.
2. Adjust, do not re-run
This is the single most useful thing to know about refining an AI edit. Asking for something again applies it a second time on top of the first, which doubles the effects rather than making them stronger. Asking for it to be adjusted changes what is already applied. The phrasing difference is small and the outcome difference is large.
- 'Make it less shaky' — adjusts the applied treatment.
- 'Add the shake again' — applies a second layer on top.
- 'Turn the intensity up' — changes what is there.
- 'Make the zooms softer' / 'make the captions bigger' — same principle, any applied setting.
If a look is nearly right, always phrase the correction as an adjustment. Re-running a pass is almost never what you want.
3. Describe the gap, not the request
When something lands wrong, the instinct is to repeat the original instruction more loudly. It works better to describe what is wrong with the result you got. 'The zoom is a beat late' gives the editor something actionable; 'add a zoom there' just repeats a request that already produced the wrong outcome.
- Say what is wrong: too early, too strong, the wrong clip, the wrong word.
- Point at the specific moment by timestamp or by what was said.
- Name the direction of the fix — softer, later, shorter, quieter.
- One correction at a time when you are being precise.
4. When to undo instead
Undo is the right move when a whole reply went in the wrong direction. Everything done in a single reply is grouped into one undo, so a single press takes back the entire change rather than unpicking it operation by operation. Passes applied from the quick edit rail also each carry their own revert, so you can remove one treatment while keeping everything else.
- One undo takes back everything from the last reply, not just the last operation.
- An applied pass can be reverted on its own from its card.
- Prefer a correction when the result is close, and undo when it is not.
5. A correction library
These cover most of what goes slightly wrong in practice. They are worth reading once, because knowing the shape of a good correction is what stops a small problem turning into a rebuild.
- 'Put back the pause at 01:14.' — restore something the pacing pass took.
- 'Make the zooms softer and less frequent.' — adjust motion intensity.
- 'Move the zoom so it lands on the word free.' — retime an emphasis.
- 'Use a clean caption style instead.' — replace the whole caption look.
- 'Make the captions bigger.' — size only, look unchanged.
- 'Show fewer words at a time.' — phrasing, not styling.
- 'The music is too loud under my voice.' — adjust the level relationship.
- 'Take the shake off entirely.' — remove one treatment.
- 'That cut is a beat early — start it after I finish the sentence.'
- 'Keep the edit but make it 30 seconds shorter.'
✖ Repeating the original request when the result is wrong
Why it fails: The same instruction produced the wrong outcome once and will tend to produce it again, or stack a second copy on top.
✔ Better approach: Describe what is wrong with the result instead.
✖ Undoing a mostly-good edit to fix one detail
Why it fails: You lose all the correct work along with the one thing you disliked.
✔ Better approach: Ask for the specific change. Everything else stays.
✖ Running a pass again to strengthen it
Why it fails: A second run layers the effects instead of intensifying them, which looks noisy rather than stronger.
✔ Better approach: Ask for the intensity to be turned up on what is already applied.
Execute this workflow automatically in Cubix
Get a draft in seconds, then refine it one sentence at a time until it is exactly right.
Frequently Asked Questions
Common questions around this editing workflow.