The model gets the first answer right most of the time. Not always. The interesting question is what to do when it doesn't. The wrong answer is to retype your prompt with more words. The right answer is the R key.
What R does
When the panel is open, R opens a small input below the suggested command. Whatever you type goes back to the model along with the original task and the current command — so the model knows what it just suggested and what you didn't like. The result replaces the panel; you can refine again, or press Enter to run.
Two cases where you'll reach for it
"Too aggressive"
Your prompt was "clean up old logs" and the result was rm -rf /var/log/*. Don't accept and don't fight; refine:
R: only files older than 30 days, only .log and .log.gz, dry-run first→ find /var/log -type f \( -name '*.log' -o -name '*.log.gz' \) -mtime +30 -print"Too generic"
Your prompt was "check disk usage" and the result was df -h. You meant the deepest space hogs:
R: top 20 directories by size, recursive, human-readable→ du -hx --max-depth=4 / 2>/dev/null | sort -hr | head -20When not to refine
If the suggested command does the wrong thing in spirit — wrong tool entirely, wrong destination — start over. Refinement is for tightening, not for correcting category errors. Pressing Esc and starting fresh is one keystroke.
See also
- Find large files — refinement is part of how you tame
findqueries. - Cleanup old logs and backups — refinement is the safe way to dial in destructive commands.