The blank box, already filled in. You edit. You approve.

Deep dive · 01

Caption the photo, don't describe it from memory

In the photo lightbox, a 'suggest caption' button sends the actual image to Claude vision along with the photo's category — before, after, damage, or reference. It comes back with one short, customer-facing caption grounded in what's literally in frame, not a generic 'looking great!'. It drops into the caption field as a draft. You read it, fix a word if you want, and it saves through the same updateCaption path you'd use typing by hand. If a tech already wrote a caption, the suggestion factors it in instead of clobbering it.

• Vision-grounded — the model sees the photo, so the caption matches the shot
• Category-aware: a damage photo gets a documentation tone, an after shot gets a showcase tone
• Lands as editable draft text — your save commits it, nothing posts on its own
• Respects an existing caption rather than overwriting it

Deep dive · 03

Reply to the review without staring at it

Each review gets a suggested on-brand reply, drafted on demand. The tone adapts to the rating — a warm thank-you for a 5-star, a measured, take-it-offline tone for a 1- or 2-star. The model is explicitly told to keep it generic and gracious when the review gives no specifics rather than fabricating details. Critically, the draft is draft-only: it never writes to the review on its own and never auto-posts anywhere. It sits in an editable box until you read it, adjust it, and click reply yourself.

• One suggested reply per review, tone-matched to the star rating
• No fabricated details — stays warm and generic when the review is vague
• Draft-only: nothing is recorded until you click reply
• Edit freely before sending — the AI proposes, you're the one who replies

Draft an empty service description

An empty catalog item gets a 'draft description' button.

AI fills 2-3 sentences in the shop's tone; you trim and save. No invented prices or warranties.