How a Habr Author Turned Seven n8n Scenarios into an Autonomous AI News System

The story demonstrates how a fragile no-code assembly can grow into a full-fledged editorial system within weeks: the author abandoned seven disparate n8n scenarios and built a single Python pipeline that collects, filters, translates, formats, and publishes AI news almost without manual intervention. The starting point was classic automation chaos. Each n8n workflow handled a separate stage: news collection, deduplication, translation, image preparation, and publication.

Formally everything worked, but any error turned into a long debugging session: one scenario could complete "successfully" yet produce nothing, which would break the entire chain. At some point, the author realized the main problem wasn't individual functions, but architecture: seven processes meant seven failure points and almost zero observability. He rebuilt the new system version as a single project.

According to him, AI tools accelerated the transition: Cursor helped configure the home mini-PC and local infrastructure, Perplexity assisted with architectural decisions, and Claude Code helped convert the set of workflows into Python and deploy a containerized service. This created a pipeline where 11 workers each handle one task. The basic chain works like this: collector gathers data from RSS and APIs of roughly 160 sources every 15 minutes, scraper pulls full text, deduplicator removes semantic duplicates, ai_filter evaluates relevance, translator converts materials to Russian, llm_editor brings texts to the required style, image_worker finds or generates illustrations, and publisher releases posts on schedule.

Additionally, background processes run for GitHub repository search, video handling, and manual moderation. Special emphasis was placed on local models. Instead of ongoing expenses for cloud translation APIs, the author deployed this task on a home mini-PC running Qwen 3.

5 through llama.cpp and optimized for Vulkan. This solved two problems at once: reduced processing costs for large volumes of news and kept data within private infrastructure.

By current numbers, the system has accumulated 7127 records in the database, and the weekly incoming flow reaches approximately 1000 materials. But the output isn't an endless feed: after filtering by relevance_score, editorial review, and slot distribution, only eight posts are published daily. This approach, according to the author, proved more effective than the previous 28 publications per day, when the audience saw too much noise and ignored most of the channel.

Another important layer is agent orchestration. Publishing decisions and operational settings were moved to OpenClaw, where a coordinator and several AI agents work: an editor, pipeline manager, source auditor, and analyst. The coordinator can simply write a Telegram command like a request to lower the daily publication limit, and it independently accesses the API, changes configuration, and confirms the result.

A full-fledged dashboard grew around this: it shows worker statuses, how each news item progresses through stages, source efficiency, queues, slot schedules, and published materials. Such transparency allows quick understanding of where content is stuck and which sources should be disabled. For instance, in the first week alone it became clear that about 20 sources provide no useful signal, and pass rate helps distinguish noisy platforms from truly valuable ones.

The main takeaway from this story is not that no-code is bad and Python is good. Rather, it exemplifies how rapidly the barrier to entry for complex AI systems has changed. The author doesn't write code himself, but managed to assemble a working product by setting tasks, verifying solutions, and managing architecture.

A combination of AI coding, local LLMs, inexpensive home infrastructure, and clear orchestration transforms a personal news project into an almost autonomous editorial operation. For independent media and niche channels, this is a signal: now value is created not by automation itself, but by quality of selection, system observability, and the ability to assemble a resilient pipeline from available tools.