— Loading
Back to all case studies
— AI Content·Bali (R&D)·February 26, 2026·9 days
SkynetLabs internal R&D
Built a 200-asset carousel pipeline that turns 1 brief into a full week of content in under 9 days from null.
— Manual content QA / day
Before: 3 hrs
After: 0
-100%
— Time per 200-slide batch
Before: ~14 hrs
After: ~45 min
-95%
— AI-tell phrases shipping
Before: ~9/batch
After: 0
linted out
— Time to first batch
Before: n/a
After: 9 days
null to ship
— The problem
What was actually broken.
Producing voice-locked carousel content at scale for client launches was eating 3+ hours per day of manual layout work in Figma. Worse, every batch was at risk of leaking AI-tell phrases ('In today's fast-paced world…') because the QA was a human reading 200 slides in a row.
The brief — set internally — was to turn one topic brief into 200 PNG-ready carousel slides, with image rotation, AI-tell linting, and a render pipeline that didn't depend on a designer being awake.
— What we built
The solution stack.
— Tools
n8n (orchestration)html2canvas (render)GPT-4o (copy)Sharp (image proc)Custom AI-tell linter (60+ phrase list)
— Implementation (9 days)
- →Day 1-2: Wrote the 60-phrase AI-tell linter against 500 hand-rated slides.
- →Day 3-4: Built the n8n orchestrator: brief → outline → slide copy → layout JSON.
- →Day 5-6: html2canvas headless render with brand-token CSS injection.
- →Day 7: Image rotation log so the same stock asset never re-renders within a 30-day window.
- →Day 8: End-to-end dry run on a 200-slide batch.
- →Day 9: Production cutover, retired the manual Figma flow.
— Results
What changed.
Most 'AI content engines' you see on Twitter are a single GPT call dumped into Buffer. Ours is the opposite — most of the value is in the linter and the image-rotation log, not the LLM call. The LLM is the cheapest part of the system.
The 60-phrase AI-tell list was built by hand-rating 500 slides against the Wikipedia 'Signs of AI writing' guide. Phrases like 'In today's fast-paced world', 'Let's dive in', 'It's important to note' — every one of them is a free signal to the reader that they're reading machine output. Stripping them is a one-pass regex.
The image-rotation log is the part nobody talks about. Carousels lose engagement fast when the same illustration shows up across three consecutive batches. The log tracks every asset used per day and prevents re-render within a 30-day window. Engagement stayed flat across 12 weeks of daily output — exactly what we wanted.
“This is the system every client launch now runs on. We don't talk about it on sales calls — it's the moat.”
Tools & services used
— Same builder · same playbook
Want a result like this?
Send the brief. Fixed-price scope back in 48 hours. No retainer hostage situation.
Apply for a discovery call