the result
4.3 million views converted almost 14× worse than a 122,000-view AI experiment.
That conclusion took 60 reels, 6.6 million views, 60 sets of Insights, 59 speech transcripts and one very patient spreadsheet. ChatGPT Work assembled the dataset end to end.
The important part is not that I installed five tools and glued them together. I mostly didn't. I told ChatGPT Work the outcome, pointed it at my signed-in browser and Google Sheet, and let it discover or install the missing pieces in isolated local environments. When authentication was needed, it opened a local prompt and I entered the credentials there — never in chat.
the pipeline
One request, six jobs
Browse
Open the reel grid, identify only new shortcodes, and collect the visible metadata.
Inspect
Open each reel’s Insights and capture the metrics that views alone hide.
Download
Save each MP4 locally with a conservative, sequential downloader.
Transcribe
Run Parakeet locally, exactly one video at a time, and keep the data private.
Catalogue
Write covers, transcripts, hooks, classifications and conversion metrics to Sheets.
Explain
Separate content that gets distributed from content that earns the follow.
step 1–2
Browse the account and capture the Insights
The browser pass deduplicated reels by shortcode, saved each cover and opened View Insights. It recorded views, follows, likes, comments, shares, saves, reach, engaged accounts and profile activity — the data required to distinguish entertainment from follower growth.
Tell ChatGPT Work the outcome
This is the reusable version of the request. Change the account and destination sheet, then let the agent work out the implementation. The safety constraints matter: use the signed-in browser, keep credentials out of chat, download sequentially and transcribe one video at a time.
I want you to build and maintain an Instagram Reel content tracker for my account. Use my existing signed-in Instagram session and Google Sheet. Work incrementally: compare Reel URLs or shortcodes first, then add only reels that are not already in the tracker. For every new reel: 1. Capture the reel URL, shortcode, posted date, caption, current view count and a clear working title. 2. Save the cover as a local PNG or JPEG and embed the actual image in the sheet — do not use an externally hosted IMAGE formula. 3. Open View Insights and record views, follows, likes, comments, shares, saves, accounts reached, accounts engaged, interactions and profile activity. 4. Calculate followers per 1,000 views. 5. Download the video locally. Install an isolated copy of yt-dlp if needed. Do not extract my browser cookies. 6. Transcribe locally with Parakeet MLX. Process exactly one video at a time so the computer does not run out of memory. 7. Store the transcript, transcript status, a consistent content type and the actual opening hook in the sheet. 8. Verify the new rows and report any reels whose Insights or audio could not be retrieved. Once the catalogue is complete, analyse the spreadsheet only. Compare reach, follower conversion, shares per 1,000, saves per 1,000, content type and hook. Separate content that earns broad distribution from content that gives viewers a reason to follow.
You probably do not need to install anything yourself
ChatGPT Work can create isolated Python environments, install the downloader and transcription package, check the media, and write the helper scripts. You only need to intervene for an authentication prompt, a CAPTCHA or a decision that changes external access.
If you prefer to understand the stack, the optional setup below reproduces the local pieces. Parakeet MLX is ideal on Apple silicon; FFmpeg inspects and extracts audio; yt-dlp downloads public reel media. Instaloader was useful for an authenticated profile pass, but its direct-reel endpoint was unreliable during this run.
# Optional: let ChatGPT Work run these for you python3 -m venv ~/.local/share/reel-lab ~/.local/share/reel-lab/bin/python -m pip install --upgrade pip yt-dlp instaloader # Local Apple-silicon transcription python3 -m venv ~/.local/share/parakeet-mlx ~/.local/share/parakeet-mlx/bin/python -m pip install --upgrade pip parakeet-mlx # macOS media inspection / audio extraction brew install ffmpeg
step 3–5
Download, transcribe and structure
Every video was saved into a dated local folder. Parakeet produced 55 usable speech transcripts; four reels had no detected speech and one had no audio stream. Those limitations were recorded rather than hidden.
Process one transcript at a time.
Passing the entire directory to the model at once spiked memory. A simple sequential helper made the workflow safe and repeatable.
what the data said
Viral and follow-worthy are different products
The reach engine
Instantly recognisable workplace pain is easy to send to a colleague. The joke is complete after watching, so the viewer has little reason to follow the creator.
The conversion engine
A specific experiment creates novelty, proof and an unfinished story. The viewer follows because they expect another useful or surprising test tomorrow.
Saves were the strongest conversion signal
Spearman correlation with followers per 1,000 views. Association is not causation, but the pattern is useful: content people want to keep is much closer to content people follow for.
The reel blueprint
- 1Start with a familiar conflict
A claim or workplace pain that makes sense before the viewer knows who you are.
- 2Make it your experiment
I built, tested or made the agents do something — not generic AI news.
- 3Show the proof
The tool, result, spreadsheet or failure should be visible, not merely described.
- 4Leave one loop open
Tell viewers what you will test next so following has a concrete payoff.
The reel I would make about this
The hook contains the scale, the automation spectacle and the counter-intuitive result in one breath. The middle proves the process visually. The ending converts because it promises the next twelve experiments, not merely a generic “follow for more.”
[0–4s — hook] I gave ChatGPT Work my Instagram account. It watched 60 of my reels, downloaded every video, transcribed them locally — and then told me why 4.3 million views barely converted. [4–11s — show the profile grid] I gave it one spreadsheet and my signed-in browser. It opened every reel, grabbed the cover, caption, view count and the full Insights. [11–19s — show View Insights] That meant follows, shares, saves, reach — everything Instagram normally makes you inspect one video at a time. [19–27s — show downloads / transcription] Then it installed its own downloader, saved all 60 videos, and ran a local transcription model over them. One video at a time, because I nearly cooked my computer the first time. [27–36s — reveal the sheet] It turned the whole account into this: searchable transcripts, hooks, content types, covers and follower conversion for every reel. [36–48s — the result] And this was the surprise. My 4.3-million-view workplace joke earned 0.36 followers per thousand views. This 122-thousand-view AI-agent experiment earned 4.87. Nearly fourteen times better. [48–57s — takeaway + CTA] Funny, relatable videos get shared. Specific experiments make people want to see what you build next. I put the full workflow, prompts and tools in the guide in my bio. Follow — because I’m testing the blueprint on the next twelve reels.
The next experiment
Publish five concrete AI experiments, three workplace sketches, two useful AI tutorials, one personal story and one internet discovery. Measure reach, saves and followers per 1,000 separately. Then run this workflow again.
