The problem
Posting every day for 30 days takes hours of writing and design. Fully automatic posting is risky: AI text sounds generic, and one bad post goes out under your name. I wanted the speed of automation with my voice and my final say.
AI case study · Human in the loop · Personal tool
An AI that writes like me, and waits for my yes
A command-line agent that drafts LinkedIn posts in my own voice, makes the images and slide decks to go with them, and publishes on a schedule, but only the posts I’ve approved. It runs my 30-day “AI Terminologies Explained” series.
- Role
- Backend Developer
- Models
- Gemini (text) · Flux (images)
- Timeline
- July — November 2026
From post history to published, with me in the middle
- Learn voicePast posts → style profile
- DraftPost + image + deck
- I approveAPPROVED marker, or nothing posts
- ScheduleChecks every 15 min, Lagos time
- Publish oncePOSTED marker stops repeats
Decisions worth explaining
Voice, not template
Learned from my own posts
It reads my LinkedIn data export safely, removes duplicate posts, and has Gemini describe my style. That profile goes into every prompt.
Approval gate
Nothing ships without a yes
A scheduler walks a dated queue. It posts only approved items, marks them posted so they never go twice, uses different times on weekdays and weekends, and gives up after three failures.
Clean output
Reads like a person wrote it
Filters strip the model’s “Here’s your post…” chatter and Markdown, because LinkedIn shows plain text only.
Problems I hit, and how I solved them
Problem
Learning a voice is expensive
Analysing hundreds of past posts on every run would burn tokens and time.
Solution
Analyse once, save the profile
Gemini studies my post history once and writes a style profile, which is saved to a file and reused in every prompt. My post history is also fetched once and kept locally, with duplicates removed.
Problem
Regenerating forever
It’s tempting to keep hitting “try again” until a draft feels right, and every retry costs money.
Solution
Capped attempts, cheaper model
Each post and image gets at most five attempts, and I can type feedback to steer the next one instead of starting over. Text uses Gemini’s lightweight Flash-Lite model.
Problem
Publishing shouldn’t need the AI
If posting regenerated content, what goes out wouldn’t be what I approved.
Solution
Save the approved draft, post it as is
Approved posts are saved to a dated queue. The scheduler publishes that saved text exactly as written, with no model call at posting time.
How it’s built
Python · Gemini via the google-genai SDK · Cloudflare Workers AI (Flux) for images · python-pptx for decks · HTML/CSS designs rendered to PNG with headless Chrome · my own LinkedIn OAuth and posting client with token refresh.