How the video was made: the pill and document views were recreated for it. The results come from a recorded Cue run on synthetic meeting data, edited for time.
https://t.co/zdMuEdi3Wd
The meeting ended. The follow-up still had to be written.
One request to Cue: “Create meeting notes with decisions, action items, owners, and deadlines.”
Out came a Word doc saved beside the transcript: 3 decisions, 3 action items, and 1 open question left unassigned.
Any Friday, 4 PM. Coffee’s cold. The report’s still due.
I don’t want another AI button. I want that report off my plate.
That’s the ambition behind Cue Office. We call it Coffee.
Concept artwork · Product ambition. https://t.co/LACj11IMbS
@exceldictionary The folder plan is useful; I'd also want a saved old-path → new-path manifest and a small dry run before moving the whole folder. A tidy folder is easy to judge visually; broken references are easier to miss.
@hwchase17 For a spreadsheet agent I'd want to join each run to a reopened-file check: formula still present, recomputed value correct, source link intact. Then slice failures by model and tool version. A successful tool call is a weak proxy for a usable deliverable.
@karpathy Tried the speech-recognition search. The HN comments under each video are the best part: you can see why it was cited. Two top results look like reuploads of the same Vista demo—could those be grouped while keeping both comment histories?
Monday meeting notes: “Maybe ship Friday—ask Alex.”
That’s an open question, not an assigned task.
A Cue workflow concept: separate confirmed actions, open questions, and owners to confirm.
What gets lost in your handoffs?
Concept, not a recorded demo.
For Cue Office, here's a spreadsheet test I'd run:
10 + 20 = 30.
Export. Reopen. Change 10 to 11.
Does the total become 31, with the formula intact?
A proposed check, not a test result yet. What breaks most often in your exported files?
Cue 1.4 is out: updated voice input + Agent automation.
Try one real task:
• Dictate a short draft
• Ask Agent to turn a public page into 3 takeaways
Where does it work well—and where does it get stuck?
https://t.co/zdMuEdhw6F
@thsottiaux One thing I’d love: a clear handoff view—what changed, what still needs my approval, and where to resume if a task stalls. A space feels useful when I can leave, come back, and understand the state without rereading the whole chat.
Help us make Cue better 🐛
Be the first to report a new, reproducible bug we confirm and get another month free. DM me the steps + a screenshot or recording.
One giveaway entry and one bug reward per person. Multiple-account entries are ineligible. Up to 2 months total.
I’m giving 30 people a free month of Cue 🎁
GPT-6 + Claude 5.5 included.
Speak your request. Share your screen context.
Follow + reply: what task would you love to automate?
First 30 qualifying replies get a month. Plan limits apply.
https://t.co/P27MijTInC
#AIAgents
@BrenBuilds@cursor_ai Publishing individual runs makes a speed claim much easier to inspect. I’d like to see the slowest runs and failed attempts alongside the median: those are often where an agent workflow loses its time advantage. Is your timing end-to-end, including tool calls?
A test I want every office agent to pass:
Give it a PDF with one conflicting date and one missing number.
Does it flag both, or quietly fill the gaps?
A polished answer is easy to trust. Visible uncertainty makes it easier to check.
A design priority as I build Cue.
Honored to be featured by @googlegemma in the Gemmaverse. Cue uses Gemma locally to turn speech into clean, natural text in ~488ms—without waiting on a cloud round trip. Local-first makes voice AI feel native.
@googlegemma Thank you for featuring Cue in the Gemmaverse. Running Gemma locally is a core part of making desktop voice interactions feel genuinely instant. Excited to keep building at the edge.