Qwen3.6 35B Transforms Workflows Through Skill-Based Prompting
Original: Qwen3.6 35Ba3 has changed my workflows and even how I use my computer View original →
How One User Rebuilt Their Workflow
A LocalLLaMA user shared a detailed account of how Qwen3.6 35B A3B changed not just their coding workflow but their overall computer usage — earning over 340 upvotes. The approach is less about raw model capability and more about a structured system built on top of it.
The Three-Layer Workflow
First, use Codex to perform specific tasks while documenting the process — including errors — as a reusable "skill." Second, feed that skill to the pi agent. Third, Qwen3.6 equipped with these skills reliably handles tasks that previously required significant trial and error.
Real Tasks Automated
- DevOps management on a VPS
- Converting old PDFs to EPUBs using Docling
The skill documentation acts as pre-loaded context: it tells the model what errors to expect and how to handle them before the task begins.
The Broader Point
Local LLM utility depends as much on how well users structure their knowledge systems as on model capability itself. Qwen3.6's reasoning paired with a well-maintained skill library and agent framework enables personal automation at a level that previously required much larger or proprietary models.
Related Articles
HN focused less on whether local LLMs fully replace frontier models and more on where they already make sense. The thread turned into a practical debate about Gemma, Qwen, agentic coding, memory limits, cost, and privacy.
LocalLLaMA users reacted strongly to a small but practical vLLM nightly change. The new Qwen3+ streaming parser is aimed at mid-turn stops and streaming tool-call failures that can break Qwen3.6 agent loops.
Alex Ellis’s post resonated because it framed local LLMs through business use, control, cost, and agent reliability instead of a simple benchmark ladder.