47!!
That's how many things had to happen between the moment I hit stop on a recorder and the moment something was live across all my platforms. Forty-seven manual steps. Fourteen of them were me personally moving text from one tool to another. Copy-paste, reformat, upload, schedule.
Count your own manual steps this week. It's a strangely useful exercise. Counting changes things, you start looking for the fix instead of living with the friction.
Imagine a piece of furniture sits in the wrong spot in your room. You bump into it every day. Moving it isn't hard, it just takes a bit of planning, so you keep telling yourself you'll deal with it later. Instead you get a small hit of annoyance every day and move on. It doesn't feel like a big deal until you add up a year of those moments.
Work runs the same way. We all have a process or a recurring problem that costs us a few minutes a day, and it's easier to work around it than to stop and fix the root cause. Eventually the friction adds up past the cost of just fixing it. That's exactly what happened with my content system. I stopped treating it like a creative problem and started treating it like an engineering problem.
In this email, I'm going to walk you through everything: how the system is designed, how the repurposing agent works, how clips get scored, how analytics feed back into decisions, what my actual week looks like, and the one piece that stayed broken until recently.
If you're a creator or marketer running any kind of content operation, most of this is buildable. Some of it is technical. None of it requires you to write code.
Step 1: The map
Before automating anything, you need a map.
Not a vague sense that you should be more efficient, an actual list: what platforms you publish to, what each one prefers, and how many times you touch a piece of content by hand before it reaches anyone.
Here's how I'd do the mapping exercise.
List every platform in order of strategic importance to you, not follower count, importance to your actual goals.
For each one, write down its input format: word count, aspect ratio, tone, link rules.
Trace the path your content takes from source to each platform, and write down every step. Circle the ones only you can do, creative calls, approvals.
Box the rest, the steps that are just movement, text going from A to B.
Bingo! The boxed steps are your automation targets.
My own map today:
YouTube (Silicon Valley Girl), 1.55M subscribers, the discovery engine. About 82% of the new Future Proof newsletter (yes, the one you are reading!!) subscribers come from here. This year: 7.5M views, 50.9M minutes watched, 185,563 shares. The interview is the source material, everything else is downstream from it.
Instagram: 1.3M subs, the visual brand layer. Reels and carousels, shorter attention span, higher visual bar.
TikTok: 3M followers, zero dedicated strategy, just cross-posted from the main system. I'll come back to this.
LinkedIn: almost 70K followers, 15M+ impressions a year, $80K in direct revenue last year. Professional authority, and increasingly, AI search visibility. Consistent, formatted text posts do well here.
X: podcast clips and occasional tech posts.
Future Proof, this newsletter: 21,500 subscribers, 50% open rate (thank you!). The economics of a direct email relationship are completely different from any platform. No algorithm between me and you.
The core design principle: one input triggers almost the entire stack. I record one interview or tutorial. That's the starting point for YouTube, Instagram, LinkedIn, X, TikTok, and this newsletter. I still build separate content for each platform, but starting from one source cuts my personal time roughly in half.
The system that gets one podcast episode to YouTube, IG, TikTok, LinkedIn, X, and the newsletter without redoing the work six times.
This isn't a new idea. Most people don't do it because they never build the infrastructure that makes it automatic. Once the map exists, the actual work starts underneath it.
Step 2: Skills, not prompts
After I finish an interview, the transcript goes into Trint, then lands in our content database, one row per episode. From there, different team members trigger different skills depending on what needs to get made.
A prompt tells an AI what to do once. A skill tells it how we work. The difference matters more than it sounds: a prompt gets you a draft you have to fix, a skill gets you something you can actually use.
Each skill we've built contains four things: what's previously worked for us, real examples of posts and clips that performed, not generic best practices; our best-performing text, verbatim samples for tone and sentence rhythm; who the audience actually is, what they care about, what they already know; and platform-specific rules, character limits, link policies, what performs there versus what just looks like every other account.
The actual skill file that checks every script before it ships. This is what "write like me" means when it's not a vague instruction.
Building your own:
Start with 5-10 of your best-performing posts on that platform, these become the calibration reference.
Write 3-4 sentences on your actual audience, specific enough to be useful. "AI-curious professionals skeptical of hype" works. "My audience" doesn't.
List the platform's quirks: LinkedIn penalizes external links in the body, Instagram front-loads the hook before the cut, X threads lose engagement after tweet four.
Add a banned list, words and patterns that sound like AI, or just don't sound like you. Ours has about 40 banned words and 12 structural patterns.
Each skill is owned by whoever runs that channel. When LinkedIn output goes stale, the LinkedIn lead updates the skill, not me. Build time: 2-3 hours the first time, 20 minutes to update.
I'm still not happy with the X skill. It comes out too structured, too "thought leader." We override about half of what it produces. More example posts and a tighter banned list would probably fix it. Haven't done it yet.
Clip scoring
A separate process runs on every transcript, in parallel, looking for moments worth turning into short-form video. Each passage gets scored 1-10 against four criteria: a surprising stat, an emotional moment, a counterintuitive claim, or a standalone idea that makes sense in 60-90 seconds with no context. Anything above a 7 gets flagged for my editors, edited for Reels on Instagram and Shorts on YouTube, then routed into the same Telegram queue as everything else.
The clip-scoring skill, built on 4 months of our own Instagram data, not a generic template.
The prompt is close to this: read this transcript, score each paragraph for standalone short-form potential against these four criteria, return the top five, with the exact quote, the score, the reason, and a suggested text overlay. Claude runs this in about 30 seconds per episode. What comes out is concrete enough to hand an editor with no back-and-forth about which clips to pull.
The analytics layer
Building the repurposing system gets you speed. Building the analytics layer gets you compounding improvement, because you can't steer what you can't see.
SVG Command Center is a custom dashboard on Node.js, hosted on Railway, pulling live data from YouTube, Instagram, and Beehiiv into one seven-day rolling view.
Three days after every YouTube video goes live, an automation pulls the retention data and writes a post-mortem to the database. By day three the algorithm has mostly made its distribution decision, so I know within 72 hours whether the hook worked and where people dropped off.
A custom dashboard aggregates data across all platforms and compares it against weekly goals.
Two audits run every Monday. One reads the previous week's data and surfaces what performed, with the number that proves it and the element likely responsible. The other scans the week's workflow logs for manual steps that shouldn't need a human, and flags them. Every Monday, one or two things come off my plate.
If you're starting your own layer, build for the one question you most need answered each week. For me, it's the ratio of views to new subscribers. A video with 200,000 views and 400 newsletter subscribers is worth more to my business than one with 500,000 views and 200. The first is more buildable, the second is just more watchable. Pull publish date, platform, views at D+3, D+7, D+30, follow rate, subscribers driven, and notes into one row per piece of content. Add complexity once you know what you're actually asking.
What the week looks like
Monday: two audits run overnight. I review both in about 20 minutes, then jump on a call with my team on Wednesday to talk through what's getting automated next and how things are performing.
Record day: a 60-90 minute interview. The transcript hits the database that evening, the repurposing agent fires, clip scoring runs, and everything lands in Telegram before I go to bed.
Post-record: my team opens Telegram to four drafts, LinkedIn, X, a Telegram update, a newsletter section. I tap approve or reject. If I reject, I leave a note on what was wrong. Total time: about 10 minutes.
Newsletter send day: the draft already exists from the repurposing flow. Edit for 30 minutes, finalize the subject line, send.
D+3 after a YouTube video: the retention post-mortem lands automatically. I read it. Five minutes.
Total personal time in the pipeline, on a typical one-episode week: about 2.5 hours, including the interview itself. That doesn't mean the team isn't working, editors are editing, designers are designing. The system just removes the bottleneck of me having to touch everything before it can move.
That's the system on a normal week. It has one blind spot.
This issue is brought to you by Hubspot
Where it broke
TikTok has zero dedicated strategy. Yet, we got 3M followers. Clips land there as the last stop in the flow above. That part works.
For a while, TikTok lived completely outside my analytics layer. SVG Command Center pulled YouTube, Instagram, Beehiiv. TikTok sat in a separate tab I checked irregularly, which meant the Monday audit couldn't read TikTok data at all. A clip could pull 200,000 views there and I'd have no idea unless I happened to check that tab the same week. I was publishing to 3 million followers and not looping any of it back into the system. The skill discovery audit eventually flagged this for me. That's the system working, just later than it should have.
If you don't want to post to TikTok and track the results by hand, HubSpot integrated a TikTok management tool directly in its platform.
TikTok in HubSpot lets you schedule and publish TikTok content alongside your other channels in the same calendar, reply to comments from HubSpot's Social Inbox, and see TikTok performance next to everything else you're already tracking, instead of reconciling five tabs yourself.
If TikTok is one of your platforms, even just for cross-posting, make sure it isn't a blind spot the way mine was.
TikTok in HubSpot. Once it's connected, TikTok sits in the same calendar and the same reporting as everything else you're already tracking
#HubspotMediaPartner
What my system still doesn't do
YouTube comments are not being fed back into the intelligence layer. I read them manually. That's a gap.
YouTube comments are honestly my main source of real improvement for the podcast.
The voice filter still misses on X about 20% of the time.
The point is that the imperfect system I have now requires roughly 2.5 hours of my time per week in the content pipeline. I can improve it incrementally.
I couldn't improve the 47-step version because I was too busy running it.
How to start
Do the mapping exercise first. Count your steps, write the number down.
Build the database before any automation. It needs somewhere to write.
Build the repurposing piece before the analytics piece. It pays you back immediately, the analytics piece pays you back over months.
Write your voice filter before your platform prompts. If the filter isn't good, your outputs stay generic.
Set up one place, Telegram or otherwise, where everything lands for approval. That's what makes it manageable instead of chaotic, not just for you, for anyone on the team touching the pipeline.
My first version generated LinkedIn posts that sounded like a press release about me. The voice filter didn't exist yet. I built it because the bad outputs taught me what it needed to catch.
Start imperfect. Measure it. Fix what breaks.
Marina 💜