1.3M+ views and 2,500+ new followers in the first five days — an autonomous system that generates on-voice posts with realistic visuals and publishes on one-tap approval.
An independent language teacher ran their Threads account by hand — writing every post, mocking up screenshots, publishing, and manually tracking what landed. It ate hours, output was inconsistent, and the account mostly sat still.
The content had to hit a very specific personal voice — sharp, ironic, dark-humoured — to earn reach on Threads, while quietly funnelling readers toward lessons. Doing that every day is a full-time job on top of an already full-time job. So it didn't happen.
The goal: turn content creation into a pipeline — generate → approve in one tap → auto-publish — without losing the author's voice, and grow both reach and inbound leads.
I built an autonomous content system that runs the entire loop.
This mattered more than the writing itself. Before generating anything, the system pulls every post from the account and ranks them by reach, identifying which formats actually perform for this audience — then leans into the winning pattern while mixing in enough variety to stay fresh. Most content automation skips this and generates blind; matching the platform's proven format is half the battle.
Raw model output always sounds like a machine. The system builds a voice profile from the author's real posts and retrieves similar high-performing examples for each generation (few-shot over vector search), then runs every draft through an LLM-as-judge pass that scores it against the author's actual voice and re-tunes until it passes — real convergence toward a specific human style, not generic tone descriptions.
The winning format is a POV screenshot — a short hook caption over a mock notification (a message from a "student," a payment alert). The system generates these to be visually indistinguishable from real ones, using the author's actual posts as references. Text and details are editable straight from chat in plain language, and a continuity feature lets the author produce "part 2" of a viral story with the same character and background.
The author reviews finished cards in Telegram — approve, edit, reject, or regenerate — and approved posts publish to Threads automatically, on schedule, via the Threads Graph API.
The author rates outputs ("their voice / not their voice"), and posts that drove a lesson enquiry get tagged — so the system continuously sharpens both the voice match and the formats that actually convert.
The teacher got a content studio in their pocket. The system proposes finished posts in their voice with realistic visuals; publishing happens on a single tap. Manual work dropped to review only.
The top post became the most-viewed on the entire account and drew a comment from one of the country's largest telecom brands — visibility this account could never have sustained by hand.
| Backend | Python · FastAPI · Celery |
| Data | PostgreSQL + pgvector · Redis |
| LLM layer | Provider-agnostic — OpenAI · xAI Grok · Anthropic |
| Images | Reference-based image model · render-template fallback |
| Integrations | Threads Graph API · Telegram (aiogram) |
| Infrastructure | Docker · dedicated server · automated migrations |