From Friday Deep-Dive · Friday, October 9, 2026 · 11 min read
One interview in, a finished episode out, and then an app to repeat it
Two phases. First, Claude edited one real episode start to finish: it transcribed the interview, wrote an editor's critique, had ElevenLabs record a broadcast-voice sponsor read, assembled the cut with free open-source video tools, placed game photos over the long answers, and produced a thumbnail plus YouTube title, description and chapters. Second, it wrote a Python web app, pushed it to a private repository and deployed it, so the sports editor pastes the announcement email he already sends, drops in the photos and video, reviews, and approves. The app is not proven live yet: two runs failed, both on account settings rather than anything to do with video. In the full post: all four prompts, the corrections that became the app's permanent rules, the real costs with one of my own assumptions corrected, and the student-data question nobody raised.
Every week our sports editor records an Athlete of the Week interview on a phone. Somebody edits it by hand, adds a sponsor message, and posts it to YouTube.
Two weeks ago I handed Claude one raw interview, 10 minutes and 56 seconds of two people talking, along with eight game photos, our logo, and the sponsor spot we had been running. I asked it to make the whole thing look professionally produced.
What came back was a finished 5 minute 26 second video: cold open, animated title, sponsor lead-in with a broadcast-quality narrator, the interview cut between wide and tight framing with a name bar and game photos over the long answers, a quick-question round, and the sponsor's logo at the end. Plus a thumbnail, a YouTube title, a description, and chapter markers.
Then I asked it to build a web app so our sports editor can do that himself every week without me.
The honest headline first: the app is built and deployed, and it has not yet finished a single live episode. Two live runs failed. Both failures were fixed. The third had not been tried when I wrote this. I am telling you about it anyway, because the first half of this works today for anybody, and because the reasons it broke are the most useful thing here.
Phase one: one episode, by hand
This is the part you can do this week. No code, no hosting, no developer.
I put the raw interview, the photos, our logo, and the old sponsor spot in one folder on my computer and connected that folder to Claude in the desktop app. Then I asked for a critique before anything else.
Every week, we record a video for our Athlete of the week. Generally it is now more than 2-3 minutes long. We then edit it and post it to youtube. After introducing the video, we have a 10-15 second sponsor message included and then dive into the interview including a q and a at the end. Then at the end we post the sponsors logo. I need to automate this process and make it look more professionally done. in the folder is the raw video for our last AOW. Review the video and offer recommendations, as a master video designer and editor, on how to make this look better and automate the process.
Note the last four words of the third sentence: "as a master video designer and editor." That framing is doing real work. It sampled frames out of the video, transcribed the audio, and wrote me an actual editor's critique with recommendations.
Then I told it to go make the thing.
The sponsor lead-in for her video is here: [PATH TO THE SPONSOR MP4] I want you to come up with the entire professionally done video, including maybe changing the voice for the sponsor ad to a more professional sounding voice, using the elevenlabs mcp I have you connected to. Also, spin up a high-quality, new video.
It picked a deep broadcast voice from the ElevenLabs library, recorded the sponsor narration, composed an instrumental music bed, and assembled the whole edit with ffmpeg, which is free open-source video software. It leveled the audio to the loudness most video creators aim for, checked its own frames, and fixed what it found.
The idea worth stealing
Here is the part I did not expect.
I gave it three corrections on that first video. The narrator mispronounced the athlete's last name. The athlete's face was on screen the entire time. And the cutting was choppy with too much zooming in and out.
Then I did something slightly different with the last one:
I want to ONLY put this in the memory notes for the next one. I do not mind the choppiness of the interview, but at times it is way too choppy. Also, the zoooming in and out needs to be a little bit less.
Typos and all. Note what it does: it tells Claude to remember the note for next time instead of re-rendering now.
Every one of those three corrections became a permanent rule in the app I built afterward. "Too choppy" turned into: only cut an answer at a pause longer than one second, never leave a piece shorter than two seconds, and hold each framing at least nine seconds. "Her face the whole time" turned into game photos placed over the long answers. The pronunciation fix turned into a "how to say it" field that the app remembers for every athlete.
Your corrections are the specification. Edit one real episode alongside the AI, write down every single thing you told it to change, and you have just written the requirements document for the tool. That is true whether or not you ever build the tool.
Phase two: the app
I asked for it in plain language, including what I did and did not want saved.
I would like to make a python project in claude code for this. I would like to upload the raw video of the interview. Upload the logo for the business along with a link of their website so you can go out and grab photos/information about their business and keep the same "Proud to support" message you had for [sponsor]. Also, have a spot where we upload 5-10 photos of the player to fill in like we just did. Then you take over to make it. If we would OAuth it and not use an api key, that would be best, however, if needed, we could use a key. I currently use Railway, however, I also use others if databases are needed at all, we wouldn't need to save the raw interviews, however, I'd like to save the business lead-ins as sometimes they are the same within the sports season.
Before writing anything, Claude asked me three questions: where should the code live, how should staff log in, and do you want to approve the scripts it writes. New private repository, one shared password, and yes, approve with the ability to edit.
It wrote the app in Python, pushed it to a private GitHub repository, and I deployed it on Railway, which is a paid hosting service. The app calls two outside services: Anthropic's API for the thinking, and ElevenLabs for the voice.
The smartest thing in the whole build was not technical. Our sports editor already emails the sales staff every week announcing who was picked. I gave Claude seven of those real emails and told it to read them. Now the editor pastes that same email into the app and the athlete, team, sport, grade, parents, highlight, and stat boxes fill themselves in. He is not learning a new process. He is pasting an email he was already writing.
Build it around the paperwork your staff already produces, not around a new form you want them to fill out.
Here is his week now.
1 Sign in and start a new episode.
Shared password. One button.
2 Paste the announcement email.
The fields fill themselves. He checks them.
3 Confirm the sponsor.
It defaults to last week's, because in a season it usually is last week's.
4 Drop in the photos and the video.
One box for all of it, five to ten photos plus the interview.
5 Review, then approve and render.
A few minutes later he gets the sponsor script, the narration, the interview topics it chose, and dots showing where it thinks the athlete is in each photo. He edits anything he wants, approves, and downloads the video, thumbnail, and YouTube text. He posts to YouTube himself.
What broke
Quite a lot, and the pattern matters more than the list.
Getting files in and out was harder than editing them. Claude could not reach our network drive, so I exported to a local folder. ElevenLabs stores its audio on a host Claude's workspace is blocked from, so the files came down through the desktop app's browser pane into my Downloads instead. Then the video files it copied back to my computer arrived corrupted, and it solved that by padding each file with 16 bytes before sending and stripping the padding on arrival, moving the big ones in 20 megabyte pieces. I did not ask for any of that. It diagnosed and worked around all three on its own.
The editing problems were caught by the machine. The judgment problems were caught by me. It found its own timeline bugs, a volume-pumping error, and music stings covering the athlete's words. It did not notice that the athlete's face was on screen for five straight minutes, or that the cutting was too choppy. A person has to watch the video.
Then the live runs failed, and not for any reason involving video. The first died on a 401 Unauthorized from ElevenLabs: my API key did not have the right permissions switched on. The second died on an error from the Claude API, because the app was demanding a response format that model would not give. Both are now fixed, and Claude added a "check API keys" button and plain-English error messages.
So: the AI wrote a working video editor. What broke was account permissions and settings. Budget your time accordingly, because that is the opposite of what you would expect.
One design failure worth naming. The first version worked for one sport. Most of our sports report times, places, or marks instead of scores, and two of ours are co-ops with different team names. It also found the athlete in photos by reading a jersey number, which does not help for wrestling, track, or cross country. All of that took follow-up prompts to fix, and it is the kind of thing only somebody who knows the beat would catch.
What it costs
My Claude subscription covered phase one and building the app. Beyond that:
The Anthropic API is pay-per-use and completely separate from a Claude subscription. The app makes three or four calls per episode. My estimate is under a dollar an episode, and I have not seen a full month's bill yet.
ElevenLabs needs a paid plan, and this is worth knowing before you start: the free tier does not grant commercial rights to what it produces. Paid plans start low, a few dollars a month, and climb with volume. One more thing to check: AI-generated music can carry a separate license requirement for advertising use, and a sponsor message is advertising. Read those terms before a sponsor spot goes out with generated music under it.
Railway, the hosting, has a correction I want to flag because I had it wrong. The monthly plan fee is a credit, not a ceiling. Usage is metered on top of it, roughly by processor and memory per month, and rendering video is processor-heavy by nature. A plan advertised at five dollars a month will not be a five dollar bill if you are rendering a five-minute show every week. Check it after the first real month.
GitHub was free for a private repository. ffmpeg, OpenCV, and the Python pieces are free and open source.
The honest answer on total cost is that I will know in a month and I will tell you.
The limits
It is not proven live yet. Said again because it matters. The full pipeline has only run end to end in test mode with stand-in answers.
Names. The narrator will mispronounce unusual last names until a person types how to say them. Somebody has to listen to the audio every week.
The sponsor script. Claude writes it from the sponsor's website and is told to use only what it finds there. A person approves every script, every time. A thin website gives you a thin script.
Finding the athlete in photos. Face matching handles clear shots. Helmets, wrestling headgear, pile-ups, and bad lighting defeat it. When it is unsure it stays wide and safe, and the editor can tap to fix it.
Editorial judgment. It picks the answers it thinks are best and drops dated references. It can absolutely cut the quote a parent would have framed. The review page lists every topic so a person can put it back.
Garbage in, garbage out. Vertical phone video that is out of focus with bad audio comes out the other end as a nicer-looking version of exactly that.
It does not post to YouTube. Somebody downloads three files and uploads them.
The question I had not thought about
Writing this up raised something I want to put in front of you, because I do not think it has an obvious answer.
That announcement email contains a high school student's name, grade, and parents' names. The interview contains the athlete's voice and face. All of it now passes through two outside companies' systems.
A photo release for publication is not the same thing as consent to send a minor's name and voice to an AI vendor. And the app does not actually need the parents' names to make the video, which means that particular detail is traveling for no reason.
I am not telling you the answer. I am telling you to decide it on purpose rather than discover it later. For us, that probably means stripping what the edit does not need and adding a line to the release. Your situation may be different, and it is worth five minutes with whoever handles your releases.
What I would tell another publisher
Have Claude edit one real episode with you first, and watch it closely. The three corrections I gave on that one video became the rules the tool follows every week. You cannot skip that step and specify it from a chair.
Build around the material your staff already produces, so the weekly job is paste, drop, approve.
And budget a few rounds of live debugging. API keys and model settings broke our first runs, not the editing.
If you only do one thing: take last week's interview, put it in a folder, and ask for the critique. That prompt is at the top of this piece, and it costs you twenty minutes to find out what your own footage could look like.
Have a good weekend.
Try this and tell me what happened.
Did it work? Hit reply. Did it come back sounding like a stranger? Hit reply harder. That’s the interesting failure, and it’s almost always fixable.
Get the next one in your inbox
Tuesday brief + Friday deep-dive. Free. Unsubscribe anytime.
Subscribe — free, twice weekly← Back to I automated our Athlete of the Week video. Here is the honest version.