The latest articles on DEV Community by Mattia Maggioni (@mattiamaggioni). https://dev.to/mattiamaggioni https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4208478%2F68d0364b-bef2-4dce-adb2-287d22a9c0f8.jpg https://dev.to/mattiamaggioni en Mattia Maggioni Mon, 26 Oct 2026 10:56:17 +0000 https://dev.to/mattiamaggioni/our-app-runs-its-own-instagram-accounts-how-the-autopilot-works-and-week-1s-real-numbers-264d https://dev.to/mattiamaggioni/our-app-runs-its-own-instagram-accounts-how-the-autopilot-works-and-week-1s-real-numbers-264d <p>I'm building an app that creates, edits and posts short videos on its own. To test it in public, it runs two Instagram accounts by itself: one in English since October 21 and one in Italian since October 22. Nobody edits the videos; when something comes out wrong we fix the app. Here's how the autopilot works, and what week 1 looked like.</p> <h2> How the autopilot works </h2> <p><strong>1. A plan per calendar week.</strong> Every Monday, in the audience's time zone, a worker plans Monday to Sunday: the slots (at the audience's best times, up to 2 a day), a role for each slot (core format, sales or test, with fixed shares counted over the last 4 weeks), the format for each slot, a hook and an ending, and an idea. Switch it on mid-week and the first plan only runs to Sunday, so the next one starts with a full week of data.</p> <p><strong>2. Formats are picked with Thompson sampling.</strong> For each core slot it draws a random number around each format's score, wider when the format has little data, and the highest wins. No format gets more than 60% of the core slots over 4 weeks. In week 1 there's no data, so every format is equal and the plan is mostly tests.</p> <p><strong>3. Videos are made about 30 hours before their slot.</strong> A Postgres cron job (pg_cron) checks every 10 minutes for approved slots going out within 30 hours and queues the video.</p> <p><strong>4. Making one video.</strong> A small language model writes the script, then fills a JSON scene file that follows the style's schema. The model never writes code. A fact-check pass reads the script first. A text-to-speech voice reads the lines and returns word timings, which set each scene's length and the captions. Remotion renders 1080×1920 at 30 fps inside a container that only runs when there's work. The language-model parts are cheap: title and caption about 0.1 cents per video, the fact-check about 1 cent.</p> <p><strong>5. Learning at day 7.</strong> Each video gets a score from reach, watch time and actions. Each number is divided by the median of the account's recent mature videos on the same network, in log scale, so a 300-view account and a 300,000-view account both have a "×1". Small numbers are smoothed toward the account's usual. Hits count at most ×3 and flops at least ×0.33 for their format. It needs at least 5 scored videos before it learns anything.</p> <h2> Week 1: the cold start, in numbers </h2> <p>Read on Monday, October 26. Only videos made by HypeFox; one demo video we posted ourselves on each account is left out.</p> <div class="table-wrapper-paragraph"><table> <thead> <tr> <th></th> <th>English account (Oct 21–25)</th> <th>Italian account (Oct 22–25)</th> </tr> </thead> <tbody> <tr> <td>Videos posted</td> <td>9</td> <td>7</td> </tr> <tr> <td>Views</td> <td>1,074 (49–175 per video)</td> <td>1,019 (120–232)</td> </tr> <tr> <td>Swiped away in the first 3 s (avg)</td> <td>57%</td> <td>58%</td> </tr> <tr> <td>Avg watch time</td> <td>6.1 s of ~34 s</td> <td>6.3 s of ~36 s</td> </tr> <tr> <td>Likes / saves / comments / shares</td> <td>3 / 4 / 2 / 0</td> <td>4 / 1 / 0 / 0</td> </tr> <tr> <td>Followers</td> <td>2 → 5</td> <td>1 → 4</td> </tr> <tr> <td>Signups from the account</td> <td>0</td> <td>0</td> </tr> </tbody> </table></div> <p>Point 5 is why week 1 can't be smart: no video is 7 days old yet, so there's nothing to learn from. The Monday report says it plainly: "no numbers yet". The first videos turn 7 days old on Wednesday, and next Monday's plan will be the first one built on the account's own numbers.</p> <h2> What surprised me </h2> <ul> <li>The best English video, made by the AI, was titled "An AI tool can't run your account by itself" (175 views, 44% swiped away, the best hold on that account).</li> <li>The polished demo we made ourselves was the most skipped video of the week: 99% swiped away in 3 seconds on the English account. The AI's plain explainers kept about 43 out of 100 viewers past 3 seconds.</li> <li>The two sales videos ("DM me PLAN") held worse than average and nobody sent the word.</li> <li>14 of 16 videos landed between 116 and 232 views: a new account's test audience, and nothing broke out yet.</li> </ul> <h2> What we changed (the app, never a video) </h2> <p>We rebuilt the main style on day 1 (drawn places, moving icons, a moving camera), added background music, and picked new script instructions with blind tests judged by models that didn't know which version they were reading.</p> <h2> Next </h2> <p>Week 2 is running: 14 videos per account, still without lessons. I'll post the numbers every Monday, including the first week the learning loop actually kicks in.</p> ai buildinpublic showdev automation