Reviewing AI output

Crux's core promise is that AI drafts and humans decide. Nothing AI-generated lands in your project without a click from you.

Answers wait in Ask

An Ask answer sits in the chat until you act on it. Read it, follow its Sources to verify claims, and only then use an Apply action or a Materialise button to write it into the project. If an answer is off, ask again with a sharper question — an unused answer has no side effects.

Generate proposals are editable

When you use Generate on a detail field, the proposal appears in place but is not yet saved. Edit it directly — trim, correct, rephrase — before accepting. What you accept is what gets stored, so the field always reflects your judgement, not just the model's first draft.

Revision history

Every item keeps revision history. When an apply or an edit changes an item, the previous version is preserved: compare revisions to see exactly what changed, and roll back if a change was wrong. This makes accepting AI output a low-stakes decision — you are never more than one step from the prior state. Any version can also carry a change note recording why it happened and what prompted it.

Habits that keep quality high

• Check Sources before you apply. A citation you can verify beats a paragraph that merely sounds right.
• Review by type where a skill offers it. Analyse & Label splits its output into highlights and frame items so you can accept the good ones and dismiss the rest individually, rather than taking a whole pass or none of it.
• Be specific. "Summarise pain points from the June interviews for the ops persona" beats "summarise this".
• Edit before accepting. Generate proposals are drafts; treat them that way.
• Validate anything that feeds a real decision. AI output can contain errors — the workflow exists so those errors stop with you.

Next

For which model runs your requests, how they are billed, and what to do when one fails, see AI requests and models.