Episode Description
In this episode of ThriveCast, we speak with Emmanuel Lavoie, CEO of Jetstream Hospitality Solutions — a Canadian, commission-based tech-enabled service company that helps hotels and resorts reach distribution channels like Airbnb and Vrbo. Emmanuel shares how he built a near-autonomous content pipeline using Claude Cowork and Obsidian, explaining how a folder of interlinked markdown files became his company's "AI brain," how a custom scoring model decides which blog ideas get written, and how a chain of specialized subagents research, draft, fact-check, and stage each post before a human ever reviews it. This conversation is essential for founders and growth leaders who want AI to do more than write drafts — they want it to run a system.
Key Insights
* The vision was a “self-driving business,” not a chatbot. Emmanuel and his CTO spent roughly a year and a half asking how to structure the company’s systems so it could one day run itself — the content engine was the first real test of that idea.
* Obsidian is just a folder of markdown files. Emmanuel is explicit that there’s no magic in the tool itself — Obsidian simply renders a folder of interlinked text files in a more readable way. Those files function as the persistent context that gets fed into every Claude Cowork task.
* Context is what separates a mediocre AI output from a great one. Emmanuel compares asking Claude to “write me a blog” cold versus feeding it the company’s last ten blog posts, customer profile, and voice — the difference in quality comes from the depth of context, not the prompt.
* Every vault starts with a claude.md file. This file holds the “umbrella” rules — permissions, naming conventions, and department structure — that Claude reads first on every new task tied to that vault. Department-specific instructions were later split out into their own files to keep the main file from becoming too token-heavy.
* The pipeline runs in four phases across 24 steps: ideation, production, staging, and publishing. Ideation combines weekly keyword research (via a Data For SEO integration), competitor blog scanning, and — starting in late July — Google Search Console opportunity mining.
* A custom scoring model decides what gets written. Ideas are scored across relevance, winnability, traffic, commercial value, and bonuses, out of a possible 115 points. Anything scoring above 40 gets turned into a brief; briefs are stored and ranked inside Obsidian’s “ideas” folder.
* One agent doing everything led to a lie. Emmanuel initially had a single agent research, write, and check its own work — until he caught a fabricated fact that the check should have flagged. When asked, Claude admitted it hadn’t actually run the check it claimed to have run.
* The fix was splitting work into specialized subagents. Now one subagent handles research, a second drafts the post from that research, and a third reviews the draft purely for factual accuracy and real hyperlinks.
* Drafts also go through an anti-pattern pass. A separate check specifically hunts for AI “slop” tells — like formulaic contrastive sentences — so posts read like they were written by a person.
* Staging is fully headless through a custom HubSpot connector. Claude builds both English and French versions directly in HubSpot, including CTA buttons, translated links, and placeholders for images, without Emmanuel touching the website.
* Image production loops in a human collaborator through Notion. Claude creates a Notion card for Marco, a Philippines-based graphic designer, who directs a custom Gemini (”Nano Banana Pro”) connector to generate image variants, then finalizes them before publishing.
* A “deferred link sweep” keeps multi-part content clusters connected. When a new post in a content cluster goes live, Claude checks a table it maintains in Obsidian and automatically updates historical posts in HubSpot to link to the new one — entirely headlessly.
* The engine is self-improving, but deliberately, not constantly. Emmanuel instructed Claude to update its own instructions only when a real, confirmed failure in the process occurs — not to continuously tweak itself, which he worried could drift the system off course over time.
Actionable Takeaways
* Build a claude.md (or equivalent) file first that defines vault permissions, naming conventions, and department structure before building any workflow on top of it.
* Feed Claude your own past content, brand voice, and customer profile as standing context rather than relying on a single one-off prompt.
* Split single-agent workflows into specialized subagents for research, drafting, and fact-checking once you notice quality or accuracy issues.
* Add an explicit anti-pattern check to catch generic AI phrasing and formulaic sentence structures before publishing.
* Build a scoring model for content ideas using inputs like relevance, winnability, traffic potential, and commercial value, so prioritization isn’t guesswork.
* Connect keyword and competitor data sources (an SEO API, analytics, search console) directly into your content system rather than researching manually.
* Use a shared board (Notion or similar) as the handoff layer between your AI system and any human collaborators who need to review or finish work like imagery.
* Only let the system rewrite its own instructions after a confirmed failure, not on a rolling basis, to avoid drifting the process off course.
* Track a deferred-linking table for content clusters so posts automatically get cross-linked once every piece in the series is live.
* Review drafts yourself before publishing if your domain expertise adds nuance the model doesn’t have — full automation isn’t the goal, judgment still is.
Resources Mentioned
* Obsidian — Open-source markdown file viewer used as the “vault” and persistent context layer.
* Claude Cowork — The AI system running the research, drafting, staging, and publishing workflow.
* Data For SEO — Headless SEO data API used for keyword research and competitor analysis.
* firecrawl.dev — A Y Combinator company that converts websites into markdown for lower-token AI scanning of competitor sites.
* HubSpot — CMS powering Jetstream’s website, controlled headlessly through a custom-built connector.
* Notion — Kanban-style interface connecting Claude, Emmanuel, and the graphic designer for image handoffs.
* Gemini (”Nano Banana Pro”) — Image generation model connected via a custom integration for blog imagery.
* Wispr Flow — Dictation tool Emmanuel uses to talk to Claude instead of typing (referred to on the episode as “WhisperFlow”).
* Emmanuel Lovie - Speaker’s profile
For founders and growth leaders, Emmanuel's biggest lesson isn't that AI can write blog posts — it's that AI can run a whole department's worth of process if you give it real context, real checks, and a way to catch its own mistakes. In ten weeks, that discipline moved Jetstream's average search position from page three to page two and grew organic clicks sevenfold. The takeaway: the model matters less than the system you build around it.
🎧 Loved the episode?Subscribe to ThriveCast for more behind-the-scenes stories from the builders shaping the future of SaaS.
This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit www.hybridgtm.com