CuePilot

Choosing the smallest useful job

I began with a product requirements document and a review of interview and meeting tools. The initial space included transcription, live answers, research, rehearsal, scoring and team workflows. I narrowed the first test to three moments: prepare what matters, keep it accessible while speaking, then review what was useful. Interviews and meetings gave that broad idea a concrete setting.

I made manual retrieval useful before depending on speech, kept every cue in the person’s own prepared words, and took live answer generation out of scope. Those choices let me test whether the aid was valuable in a real conversation without asking an AI feature to carry the whole experience.

Preparing material the speaker recognises

A session brings together a running order, questions, experience stories and short cues. The preparation workspace can organise notes and imported sources into proposed items, but proposed material stays out of the live pack until the person reviews it. That approval step is a product boundary, not just a settings choice: the live view should contain thoughts the speaker is ready to use.

One early implementation exposed a mismatch with that promise. It validated an exact source quotation, then displayed a model-written paraphrase as the cue. I changed how the assistant works underneath: it can organise material and place it in the right section, but the cue itself is always the person’s own words, taken verbatim from their source. The model never writes the line the speaker reads, and approval stays explicit. I also removed hard length caps from the core cue surfaces. Concise prompts are easier to scan, but the interface must never silently discard part of someone’s thought; longer material wraps or scrolls.

A sentence sent to the assistant lands as Suggested in exact wording, waiting for approval.

Only reviewed cues reach the live panel, and every cue is in the speaker’s own words.

Replacing two lists with one Run Order

The first layout kept the agenda and the cue cards in separate lists. When I was leading a conversation, that had me switching between where I was and what I wanted to say next. I brought them into a sectioned Run Order: each stage holds its own prompts, with an “Any time” area for material that does not belong to one section. Longer stories remain in a library and can be broken into draft cues for a particular session.

Covered points dim but remain visible, so progress does not erase context. Moving between sections and changing Current Focus are deliberate actions. The structure gives the conversation a shape without turning it into a script.

The Run Order at full width: sections scan by number, cues sit beneath.

Designing a live view for a glance

When a session starts, the preparation workspace gives way to a floating panel. The Run Order shows where the conversation is headed; a single Current Focus shows the point the speaker has chosen to hold onto. Questions, quick thought capture and search remain available without competing with that line. The panel can collapse to a cue pill that still shows the current thought and its place in the run.

Left Live panel: covered cues dim, Current Focus pinned, next point visible. Right Collapsed to a pill that still shows the current thought and its place in the run.

I refined the live hierarchy against the conditions in which it would be read. Cue text has four sizes, and the panel adapts so larger type retains useful look-ahead. Reviewing it against a light desktop led me to strengthen the live material and darken the light-mode accent. Progress stays visible, keyboard actions are available through a command palette, and the interface respects Reduce Motion and Reduce Transparency. These details serve the same goal: make one glance enough to regain your place.

The workspace in light and dark.

Letting listening suggest, never steer

The manual experience stands on its own. With listening enabled, CuePilot listens on the device, to both sides of the conversation or just the speaker, and surfaces their own cues as they become relevant: mention design systems, and the cues prepared about design systems come forward. It never writes anything new; every cue it surfaces is one the person wrote and approved. The person can inspect, ignore or pin it. A deterministic matcher and stability rules limit suggestion churn, while the Current Focus remains under direct user control. Transcripts are not retained.

This separates a helpful prompt from an automatic change of course. Listening can help retrieve something already prepared; it does not write an answer or decide what the speaker should say next.

What a real interview changed

I used CuePilot’s manual live view in a product design interview and informed the interviewers I was using it as my presentation aid. Several times, the cues helped me keep my place, recall a point I had prepared and avoid drifting off track. The panel gave me the right amount of context without asking for more attention.

I had prepared that pack through an open conversation with an LLM outside the app before narrowing it into cues. The in-app assistant is currently better at capturing material that is already well articulated. My next design work is a clearer route from open discussion to selecting and refining cues, then explicit approval for live use. The interview showed both what the manual aid already does well and where preparation goes next.