Reviewing AI output

Crux's core promise is that AI drafts and humans decide. This article covers the review workflow and how to get consistently good results.

The preview lifecycle

1. Streaming draft — skill output streams into a preview as it generates, so you can start reading immediately (and stop a run that's going wrong).
2. Review — read the draft in place. For structured outputs (highlights, personas, journey rows) each element is individually reviewable.
3. Edit — adjust the draft directly before committing; your edits are what get saved.
4. Commit or discard — accept to place it on the canvas, or discard with no side effects.

Nothing is overwritten silently, and discarded runs leave no residue on the board.

Why answers cite sources

Grounded answers and synthesis outputs cite the canvas nodes, files, or highlights they drew from. Follow the citation to verify the claim in context — this is the fastest quality check and the habit that keeps AI-assisted work trustworthy.

Steering quality

• Connect your evidence. Skills read upstream context; orphaned nodes generate from thin air.
• Maintain project knowledge. Decisions and constraints there prevent the AI re-litigating settled questions.
• Be specific in prompts. "Summarise pain points from the June interviews for the ops persona" beats "summarise this".
• Iterate rather than accept mediocrity. Re-run with a sharper instruction; runs are cheap relative to cleaning up a wrong commit.

AI disclaimers

Generated answers carry a disclaimer that output may contain errors. It's there for a reason — validate anything that feeds a real decision.

Continue the guide

Next: Saved & learned skills.