As developers, we love automation but content creation is still painfully manual for most teams.
I kept seeing the same problem across founders, indie hackers, and agencies:
- Writing SEO blogs takes hours
- Landing pages are repetitive
- Social content is inconsistent
- Existing AI tools feel fragmented or overcomplicated
So I decided to build a tool to solve my own workflow problem first and that project became Minineo.
This post isn’t a promo.
It’s a breakdown of what I built, what worked, what didn’t, and what I learned along the way.
The Original Problem
My goal was simple:
Go from keyword → published content with as little manual work as possible.
But most tools:
- Generate raw text only
- Don’t respect SEO structure
- Break when you try to scale
- Leave publishing & formatting to you
I wanted something developer-friendly, predictable, and automatable.
The Approach I Took
Instead of one big AI prompt, I split the system into clear stages:
- Intent-first content planning
- Strict structure enforcement
- Usage limits handled server-side
- Publishing treated as a pipeline, not a button
That design decision saved me weeks later.
Architecture (High Level)
- Next.js for the app layer
- Supabase for auth, data, and RPCs
- Server Actions for usage enforcement
- AI models only handle generation not business logic
Key rule I followed:
- Never trust the client for usage limits
- Always enforce limits via server-side RPCs
This made the system production-safe early on.
Content Generation Strategy
Instead of “write a blog about X”, the AI gets:
- Fixed heading hierarchy
- Explicit SEO constraints
- Word count bounds
- Section responsibilities
This dramatically reduced:
- Hallucinations
- Fluff
- Rewrites
The output became predictable enough to automate publishing.
Media Was Harder Than Text
Text was easy. Media wasn’t.
Problems I hit:
- Hotlinking external images (bad idea)
- WordPress rejecting uploads
- Social platforms requiring different auth models
Solution:
- Download → upload → replace URLs at publish time
- Treat media as a post-processing step, not generation
- This separation kept the core clean.
What Didn’t Work
Being honest a few mistakes:
- Trying to support every social platform early
- Overengineering UI before workflows were stable
- Letting AI decide structure (never again)
Each rollback made the system simpler and more reliable.
What Worked Surprisingly Well
- Strict prompts > “creative freedom”
- Build-in-public feedback
- Treating content like data, not prose
- Shipping small, testable pieces
Why I’m Sharing This Here
Dev.to helped me countless times while building this.
If you’re:
- Building an AI SaaS
- Automating content workflows
- Enforcing usage limits properly
- Publishing to platforms programmatically
I hope something here saves you time or mistakes.
What’s Next
Next on my list:
- Smarter media enrichment
- Better scheduling reliability
- Even stricter content validation
Still learning. Still iterating.
If you’ve built something similar or hit the same challenges, I’d genuinely love to hear:
- What broke for you?
- What scaled unexpectedly?
- What would you do differently?
Happy to answer technical questions in the comments
Top comments (2)
This is a really honest and valuable breakdown. I like how you focused on solving your own workflow first instead of chasing features or hype — that usually leads to much more reliable systems.
The decision to split generation into clear stages (intent → structure → publish) really resonates. From my experience, most AI content tools fail not because of the model, but because they treat content like free-form text instead of a pipeline with constraints.
Also appreciate the callout about never trusting the client for limits. That’s one of those lessons you only learn after something breaks in production.
The media section felt especially real — text is easy, but publishing and assets are where automation gets messy fast.
Thanks for sharing what didn’t work too. Posts like this are way more useful than “here’s my tool” stories. Curious to see how stricter validation and scheduling evolve next.
Solid approach splitting the pipeline into stages — that's where most AI content tools fall apart. They try to do everything in one shot and the output is generic mush.
The asset management pain is real too. I've been working on a tool called PostAmmo (postammo.com) that takes a similar philosophy but focused on the ideation/brainstorming side — generating viral content angles and hooks before you even start writing. The idea being that the best SEO content starts with a compelling angle, not just keyword stuffing.
Curious about your intent extraction step — are you using structured prompts or letting the model figure out intent from the raw input? That stage seems like the make-or-break for output quality.