I used to write blog posts the old way. Open a blank page. Stare at it. Write something. Rewrite it three times. Publish. Repeat every two weeks when I remembered to do it.
Now I have a system that writes and publishes for me. Not a spam bot. Not auto-generated garbage. Actual technical articles in my voice, posted on a schedule, while I'm doing something else.
Here's the honest breakdown of how it works, what it costs, and where it falls apart.
The Problem: Consistency
I've been running a few side projects for a while now — a GPU rental setup, some Telegram bots, a Garmin watch face. Every few months I'd think "I should write about this" and then... not do it. The gap between "built something cool" and "published an article about it" was basically infinite.
I tried content calendars. Reminder apps. The usual productivity hacks. None of them stuck because writing is work, and work needs time I wasn't willing to consistently allocate.
So I asked myself: what if the system didn't just remind me to write — what if it actually wrote?
The Architecture (No Hype, Just Scripts)
The setup is three components:
- Topic selection — a rotating list of drafts and ideas, picked by a simple rule (oldest first, or whatever I'm excited about)
- Writing agent — a local LLM (Qwen 3 Coder 30B on my RTX 3060) that takes a topic + my notes and produces a first draft
- Publishing agent — a cron job that posts to Dev.to via API when an article is ready
That's it. No orchestration framework, no LangChain pipeline, no vector store for "content strategy." Just scripts, cron, and an LLM.
Topic Queue (markdown files)
→ Pick next topic
→ LLM writes draft (in my voice, from my notes)
→ I review and edit (takes 10-15 min)
→ Publish via Dev.to API
→ Notify me with the link
The Writing Process
The key detail: the LLM doesn't write from scratch. It writes from my notes.
Each article topic has a markdown file with:
- What I built
- The actual architecture
- Numbers (costs, earnings, time spent)
- My honest opinion on what worked and what didn't
The LLM takes those rough notes and turns them into something readable. It's not creative — it's structural. It takes my bullet points and makes them flow.
Example: My notes for this article literally looked like:
- Problem: don't write consistently
- Solution: AI writes drafts from my notes
- Architecture: cron + LLM + Dev.to API
- Honest part: still need to review, not fully hands-off
- Cost: $0 (local LLM)
The LLM expanded that into what you're reading now. I spent 12 minutes editing it. That's the actual time savings.
What the Articles Look Like
I have a few rules the system follows:
- Personal voice — first person, casual, not corporate
- Specific numbers — real costs, real earnings, real time spent
- Honest downsides — if something is annoying, say so
- Code examples — actual working scripts, not pseudocode
- No clickbait — the title should describe the article
The result? Articles that read like I wrote them, because fundamentally I did. The LLM is just the typist.
The Schedule
Every two days, the system picks a topic and produces a draft. I get a notification with a link to review it. If I'm busy, I skip it and it moves to the next topic. If I have 15 minutes, I edit and approve.
The publishing happens automatically after approval.
In practice: about 60% of drafts get published on schedule. The other 40% I either skip because the topic feels stale, or I rewrite heavily because the LLM missed the point.
It's not fully autonomous and I don't want it to be. The human-in-the-loop is the safety rail.
The Cost
| Component | Cost |
|---|---|
| LLM (local Ollama) | $0 |
| Dev.to API | $0 |
| My time per article | ~15 min (editing) |
| My time before (full write) | ~2-3 hours |
That's roughly a 10x time savings per article. At 15 articles per month, that's 30-40 hours of writing time I don't spend.
What Actually Breaks
The LLM hallucinates my numbers. If I don't include exact figures in the notes, it'll make up something plausible-sounding. I caught it claiming "$240/month GPU earnings" when the real number was closer to $50. Always verify the numbers.
Voice drift. After a few articles, the LLM starts sounding more generic. I fix this by including snippets from previous articles in the prompt as "examples of my style."
Scheduling conflicts. If I approve two articles for the same day, the Dev.to API rejects the second (rate limiting). The system now enforces a minimum 48-hour gap between posts.
Topic exhaustion. I have maybe 20 solid topics in my queue. At one post every two days, that's 40 days of content. Then I need to actually build more things worth writing about. The system can't manufacture experiences — only document them.
Why This Works (For Me)
I'm not trying to build a content empire. I don't want to game algorithms or go viral. I just want a consistent record of what I'm building, written in my voice, published somewhere people can find it.
This system does that. It doesn't replace me — it removes the friction between having something to say and actually saying it.
The articles are still mine. The ideas are mine. The code is mine. The LLM just handles the part I find tedious: turning rough thoughts into readable prose.
Should You Build This?
If you have a bunch of half-written drafts and a notes folder full of project ideas: probably yes.
If you don't already have raw material (notes, numbers, experiences): no. An LLM can't write your blog for you if you have nothing to say. It can only help you say it faster.
The real requirement isn't technical skill — it's having something worth writing about in the first place.
The Setup in One File
If you want to try this, here's the core of it:
#!/bin/bash
# publish-article.sh
API_KEY="your_dev_to_api_key"
ARTICLE_FILE="$1"
# Read markdown content
CONTENT=$(cat "$ARTICLE_FILE")
# Extract title from frontmatter
TITLE=$(echo "$CONTENT" | grep "^title:" | sed 's/title: //' | tr -d '"')
# Post to Dev.to
curl -X POST https://dev.to/api/articles \
-H "api-key: $API_KEY" \
-H "Content-Type: application/json" \
-d "{
\"article\": {
\"title\": \"$TITLE\",
\"body_markdown\": $(echo "$CONTENT" | jq -Rs .),
\"published\": true,
\"tags\": [\"ai\", \"automation\", \"side-project\"]
}
}"
Add a cron job every 2 days, a folder of drafts, and you're 90% there.
I write about building things with AI, running hardware at home, and the occasional side project that makes a little money. If any of this is useful, feel free to follow along.
→ Custom automation setups on Fiverr
→ Follow CelebiBots on Telegram
Top comments (0)