I run five active projects. I have no team. I don't work 80-hour weeks.
The way I compete with larger teams is by building automation that compounds: systems that do work while I'm building the next thing.
Here's every piece of that stack.
The projects
Before the stack, the context:
| Project | What it is | |---------|------------| | Lunary | Astrology platform with real-time transit tracking and a 2,000+ article grimoire | | Spellcast | Social media scheduling platform (what I use to run all of this) | | Artify | Automated Instagram content pipeline | | Podify | AI podcast generator at £0.03/episode | | Content Creator | Video system with ML engagement prediction |
Five products. All active. All producing output.
Social media: Spellcast
Spellcast is the hub. It manages multiple "account sets", which are groups of social accounts that post as a unit. I have six personas running: sammii (personal), Lunary, sammii spellbound, sammii sparkle, scape², and sammii dev blog.
Each one has:
- Brand voice profiles: injected into every AI generation so content sounds like the account, not a generic AI
- Boost rules: when Lunary posts, @sammiihk auto-likes and comments on a random delay. When @sammiihk posts, @LunaryApp does the same. Cross-promotion without manual effort
- Campaign mode: batch-generate posts around a theme or product update
- MCP server: I can schedule posts by talking to Claude
Spellcast runs on a £8/month VPS and handles hundreds of posts per week automatically.
Visual content: Artify
Artify generates Instagram content for Lunary. The pipeline:
- Selects a content theme (zodiac sign, moon phase, crystal, tarot card) based on what's currently performing
- Generates an image prompt using Claude (description, visual style, colour palette)
- Creates the image with Flux
- Formats it with Remotion (text overlay, branding, correct dimensions)
- Queues it to Spellcast for scheduling
The output is a consistent visual identity that would cost £500+/month from a content agency. The actual cost is Claude API tokens and Flux generation: pennies per image.
Audio content: Podify
Covered in depth in the Podify article. The short version: Claude writes the script, Kokoro TTS narrates it locally (free), ffmpeg stitches it together. £0.03/episode. Unattended.
Video content: Content Creator
The most technically interesting piece. Content Creator generates short-form videos for TikTok and Instagram Reels, with a twist: it predicts which videos will perform before they're published.
The prediction layer:
- Analyses historical performance data (views, engagement, shares by content type)
- Scores new video concepts before generating them
- Prioritises high-predicted-performance topics in the queue
The content types that score highest for Lunary: numerology and zodiac content. The system has learned this from real data and now weights the generation queue accordingly.
The video pipeline:
- Topic is scored and queued
- Script generated by Claude
- Visuals rendered with Remotion
- Audio generated by Kokoro
- Video stitched with ffmpeg
- Queued to Spellcast for posting
The glue: MCP servers
The piece that ties everything together is Model Context Protocol. I've built MCP servers for both Lunary and Spellcast, which means I can manage both products by talking to Claude.
From a conversation with Claude, I can:
- Schedule posts, check analytics, update brand voices (Spellcast MCP)
- Query user metrics, check feature usage, pull dashboard data (Lunary MCP)
- Publish articles to Dev.to and Hashnode
- Generate content variations and score them
No dashboards. No context-switching. Just Claude with full access to my product stack.
The compounding effect
Here's what makes this work as a system rather than a collection of tools.
Each system produces output that feeds into another:
- Content Creator generates videos → Spellcast schedules them
- Artify generates images → Spellcast schedules them
- Podify generates podcast episodes → Spellcast distributes the show notes
- MCP servers let Claude orchestrate all of it
The more the systems run, the more data they generate. The more data they generate, the better the ML prediction layer gets. The better the predictions, the less manual curation I need.
My job shifts from content production to system maintenance and strategy. Which is how one person runs five projects without burning out.
The actual time cost
Maintenance of the full stack: about 2-3 hours per week. That covers:
- Reviewing scheduled content for anything that needs a human eye
- Monitoring performance metrics
- Occasional system updates or new features
Content output: 20+ posts per day across all accounts, weekly podcast episode, regular video uploads, articles twice a week.
The ratio is the point. I didn't automate to be lazy. I automated to redirect time toward the work that actually requires me.