- Home
- Skills
- Research & Analysis
- Talk Coach: A Sharper Talk After Every Rehearsal
More screenshots
Works with the AI Agents you already use
Talk Coach: A Sharper Talk After Every Rehearsal
You know the results too well to notice they went past at twice the speed of the rest, you know what your acronyms mean, and the minute you ran over felt like twenty seconds.
$9
Talk Coach: A Sharper Talk After Every Rehearsal
Example session with this skill installed
Here's the recording of my rehearsal for our department's research day on Friday (rehearsal.m4a). I get a 5-minute slot plus 2 minutes of questions, and the whole department comes: electrical, mechanical and civil engineers plus some computer scientists. My slide outline is in outline.md. Can you coach me? I want to know what to fix before Friday.
- Read your context and instructions
- Compiled the talk coach
I transcribed and timed your rehearsal on your own machine. The full report is in `talk-coach-rehearsal-2026-10-19.md`, and `rehearsal.metrics.json` is saved so the next run can be compared against this one. **The headline:** the run took **6:29, which is 1:29 (30%) over your 5-minute slot.** Almost all of that time comes before the results. They don't start until 05:05, and then they go by in 47 seconds. **What already works:** your research question at 01:13 ("can we get the same answer from the normal charge that happens every night anyway?"), concrete stakes early (4,000 bikes, 400 euros a pack, 6 hours per test), and your take-home at 05:55, which is the best line in the talk. It's just not the last thing you say. **Three things to fix before Friday:** 1. **Cut about 2:15, and don't speed up to get there.** Background and methods take 78% of the talk; the results get 13%. The report has twelve cuts with timestamps adding up to about 135 s: the greeting and "Today I'm going to talk about…", the EIS detour, "I'm going to skip some of the details because I don't have time…", the pack specs, the full feature list, and the model and temperature details. Most of what you cut makes good answers for the questions. With the cuts made, you should land at about 4:40. 2. **Slow the results down.** You give them at **214 wpm, against 126 wpm** for the rest of the talk, with about ten numbers in 47 seconds and four in one sentence at 05:05. Give three results, one number each, with a pause after each, and drop the hedge at 05:46 ("I think that's probably good enough…"). 3. **Explain SOH, ICA and DCIR the first time you use them.** SOH is never explained; "battery health" works fine, as your take-home already shows. ICA and DCIR are never expanded, and "C rate" will be new to the civil and mechanical engineers. The speech recogniser misheard several terms ("SOH" as "so", "EIS" as "AS", "R²" as "U square"). I corrected for those, and they're not counted against you. A transcript can't show your voice, volume or body language, and I didn't see your slides, so none of that is assessed. Next run-through, send me the recording and I'll compare it against this one on three targets: under 4:45 total, results at 120–140 wpm with pauses, and the terms explained.
Connects securely to your tools. The creator never sees your data.
About this skill
## What it does You can't hear your own rehearsal the way the room will. You know the results too well to notice they went past at twice the speed of the rest, you know what your acronyms mean, and the minute you ran over felt like twenty seconds. This skill makes your agent listen for you: it transcribes your rehearsal on your own machine, measures what the recording shows, and reads the talk as your audience will hear it. - **Where the minutes went.** Time against your slot, and each section's start, length, share and pace, so a rushed results section or a methods section that ate the talk shows up as numbers with timestamps. - **A cut list that adds up.** When you're over, concrete passages to cut, each with its timestamp and the seconds it saves at your own pace, totalling the overrun plus a margin. Never "speak faster". - **The talk as your audience hears it.** When the main message is first said, which terms and acronyms are never explained for this audience, how many numbers come in one breath, and whether the last sentence before the thanks is your take-home. - **Verbal habits, counted and located.** Hesitation sounds, filler phrases judged in context, sentences about the talk ("now I'll show you the results"), hedges on the main result, each with its timestamp. - **Three priorities, not thirty remarks.** Each with the evidence, the change, a drill and a target the next rehearsal can be checked against. Rerun on the next recording and compare. - **Feedback grounded in what was recorded.** Every point quotes the transcript with a timestamp. Misheard terms are corrected against your outline and never counted against you, and nothing is said about your voice or confidence that a transcript can't show. ## What a run looks like In our demo the agent got a made-up five-minute research-day talk, spoken by a computer voice, with rehearsal problems planted, plus the slide outline. It transcribed the recording locally and found all of them: 6:29 against a 5-minute slot, a results section at 214 words per minute against 126 for the rest, SOH, ICA and DCIR never explained to a mixed department audience, and three hesitations a minute. It wrote twelve timed cuts adding up to 135 seconds, three priorities with targets, and set aside ten recogniser errors ("U square" for R squared, "AS" for EIS) instead of blaming the speaker for them. ## Settings Conference talks, lightning and research-day talks, job talks and seminars, thesis defence and viva presentations, lab meetings and progress reviews, online talks. English and Spanish. ## What's included - `SKILL.md`: the ten-step method and its grounding rules - `scripts/transcribe.py`: local transcription with word timestamps (faster-whisper, mlx-whisper, openai-whisper or whisper.cpp) - `scripts/import_transcript.py`: captions (.vtt, .srt) or text instead of a recording - `scripts/delivery_metrics.py`: the measures, per section, with timestamps and rerun comparison (standard library) - `references/`: transcription setup per OS with model sizes and speeds, a delivery guide by setting, drills with targets, the feedback format - `examples/`: the demo's transcript, captions, outline and the feedback it produced ## Part of the Aurora Lab research pipeline Works on its own, on any talk. Pairs well with Aurora Lab's Q&A Prep and Poster Layout; needs neither.
It works from words and timing. A transcript can't show intonation, volume, emphasis, accent or nerves, so those are left to your own listening (it offers a short self-rating) and are never guessed. Video is judged from stills: stance, position and facing the audience, not gestures in motion. Filler counts are a lower bound, since recognisers drop some even when asked to keep them. Recognisers mishear field terms and acronyms; the agent corrects for them against your outline, but a term missing from the outline may be misread. Captions give timing per caption, so pauses inside a caption aren't measured, and plain text gives no section timing. Rules of thumb (pace, section shares, fillers per minute) are coaching guides, not norms. Transcription runs on your machine: a 6.5-minute rehearsal took 46 seconds with the default model on our Apple Silicon laptop, and an older CPU will take a few times longer. It coaches the delivery, not the science, and doesn't see your slides unless you share them as text.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before any payment.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 27 days ago
Requires
Permissions
File Scopes
Reads the recording or transcript you point to (audio, video, .vtt/.srt captions or text), your outline or slides, and the metrics of an earlier rehearsal; with a research/ folder also research/talks/ and research/claims.md. Writes the transcript (.json and .txt), the measures (.md and .json) and one feedback file, in research/talks/ when that folder exists, otherwise next to the recording. Terminal: runs the bundled Python scripts, ffmpeg to convert the audio (and, for video, to take stills into a temporary folder in the project that is deleted at the end), and a pip install of a recogniser into your project's environment only if you agree. Network: the recogniser downloads its model once on first use; the optional cloud transcription (Groq or OpenAI) uploads the recording and runs only when you ask for it. Never edits your recording, slides or outline.