She records a video, closes her laptop, and goes for a walk—or sits in a doctor's waiting room. On the way, she sends one line to a bot from her phone: 'By the time I'm back, cut out the silences, filler words, and retakes.' When she gets home, the rough cut is already waiting.
That is how Louise de Sadeleer, a marketer at video startup Tella, runs her workflow. No human editor, no quitting her day job, and no feeding viewers generic AI slop. In three months, the system brought her +6,500 Instagram followers, growth on TikTok, and a YouTube video with 50,000 views. On Greg Isenberg's podcast, she shared the whole playbook openly: 'I don't believe in gatekeeping—it's bad karma. If I'm sharing something, it means I'm already five steps ahead.'
Why an 'Always-On' Bot Beats a Regular Chat
A regular AI chat works fine at a desk, but the best ideas, tweets, and hooks hit you on the move. If every find forces you to open Notion, hunt for a table, and save a screenshot by hand, you will quit in a couple of days.
That's why Louise moved to Grok Bot—xAI's cloud agent with schedules, bot group chats, and connections to external services. Drop a link or screenshot from your phone into the chat, and the bot logs the creator in your database, grabs a tweet screenshot, and drops it into the right project's media library. The goal is simple: hand 80% of the routine to agents and keep 20% for yourself—scripts, motion design, and brand taste.
Four Weekly Research Runs and Why Daily Bots Backfire
Her first mistake was scheduling research bots to run every day. Louise dropped that fast: daily runs burn through Grok Bot limits, and the trend digest drowns in minor noise. Switching to a weekly rhythm solved both.
| Routine | Where It Pulls Data | What It Outputs |
|---|---|---|
| Outlier post finder | Creator database in Notion | Topic, hook, and gap over normal views |
| TikTok hook breakdown | Saved videos folder | Opening lines sorted by delivery style |
| Trend digest (Trend Steady) | Substack emails, X, and news | Signals that turn into trends when repeated |
| Rising keywords (YouTube Yuri) | VidIQ service | Search terms gaining momentum |
Behind the first row sits the golden rule of the system. An outlier post is an unusually viral video that gets many times more views than a creator's usual baseline. Once a week, the bot catches those spikes, yet Louise picks the creators—between 15 and 25 on her list—strictly by hand. The moment you let AI pick your taste, quality drifts off course.
In the second routine, the bot pulls the first 2 sentences from saved TikToks—the hook that grabs viewers in seconds so they don't swipe away—and sorts them by type (contrarian, relatable, questions). Louise intentionally borrows hooks from outside niches like dating and psychology. For instance, she saved a video opening with 'If a parent's hug makes you uncomfortable, it doesn't mean you're cold' and adapted its structure to B2B marketing: even B2B buyers are regular people.
Trend Steady (whose base prompt was written by Louise's friend Miranda) reads paid Substack newsletters from email, X, and news. A single mention goes into signals, while a topic popping up across multiple places—say, a16z -> Substack -> X talking about MCP as a user acquisition channel—gets promoted to a trend. YouTube Yuri pulls rising keywords from VidIQ. With zero prior YouTube experience, Louise caught the rising Remotion query, made one of the first videos on it for Tella, and hit those 50,000 views while boring product demos got ignored.
Two-Bot Group Chat and Advisors Instead of Script Generation
On Mondays, Production Polly steps in. This bot manages the content calendar in Notion and books four-hour blocks in Google Calendar for scripting and filming, so videos don't get pushed to 'tomorrow'.
Still, Grok Bot's most underrated feature is bot group chats. Louise put Production Polly (who knows her ideas, style, and hooks) and Trend Steady (who acts like a journalist) into one thread. Her command was one line: 'Combine my ideas with fresh trends.' In a single run, the bots pitched three packaged topics: a video on all connected MCPs, a breakdown of agent group chats, and a guide on avoiding AI rate limits. Together, the agents debate and sharpen each other's ideas noticeably better than a single bot alone.
At that point, automation stops. Louise writes scripts by hand, then runs them through four advisor skills—saved sets of criteria and past edits that let the agent critique a draft instead of writing from scratch. Idea Advisor tells her bluntly whether an idea has depth or only sounds cool in her head. Hook Advisor checks the intro against her hook library and templates like 'people used to do X -> in 2027 they'll do Y'. Script Advisor checks structure. And Miranda Feedback is trained on past notes from Louise's short-form video mentor to catch her usual edits early.
Without guardrails like these, bots easily run wild. Louise warns plainly: 'They're super smart, but you can't trust them blindly—sometimes they're such eager overachievers that you have to hold them back.' Skip strict skills, and out of pure helpfulness the agent will bury Notion in walls of text.
A bot owns a distinct area of work: Louise runs 4–5 work roles for content plus a separate bot for hobbies like pottery so contexts never mix. A skill is a repeatable method or strict formatting rule, like how to score a hook or how to save a concise note in Notion.
Mobile Editing via Tella MCP, Model Choices, and Honest Caveats
Creators don't burn out from spending a minute in front of the camera; they burn out hunting down tweets, screenshots, and B-roll. Louise removed that grind with Tella MCP, Tella's official cloud server with OAuth login built on MCP (Model Context Protocol, an open standard for connecting AI agents to external tools without manual file copying).
On the go, you drop a link into Grok Bot, and it uploads tweet screenshots—even Figma graphics—straight into your project's media library. After recording, Louise goes for a walk and kicks off a rough cut to strip silences and 'ums' (Tella's built-in agent also swaps layouts, adds smart zooms, and clips Shorts). At the end, YouTube Yuri reads the transcript via MCP, checks VidIQ keywords, suggests titles, writes the description, and adds chapters.
For 90% of editing tasks, Louise uses Claude Opus 5.5 at medium effort ('there's just no point in high'), saving pricey Claude Fable 5.2 strictly for heavy 3D and Blender work. Greg Isenberg pushed back: he leaves high on by default, betting that in a couple of years we'll look back amazed at how cheap tokens were.
The caveats here are honest too. Louise works at Tella, so her editing relies on Tella MCP—Greg noted he has no affiliation with the company, and you could build a similar setup with local Claude Code or Codex. Cloud agents aren't flawless yet either: right before recording the podcast, Louise lost her diagrams and admitted she had been 'fighting with Grok Bot' over missing data. Finally, there are no ready-made bot configs or public Grok Bot quota charts for video work, so you will need to tune run frequency and guardrail skills yourself.

