A developer in the Tongyi ecosystem, wsxiaoys, published a method on GitHub Gist that does something unusual: it dumps GPT-5.5 Pro's reasoning trace directly into the start of Qwen 3.8's context. The result is that Qwen's output style closely mirrors the target model.
No fine-tuning. No distillation. Just a prompt.
The original post hit 90 points and 35 comments on Hacker News, making the front page. Gist link:
https://gist.github.com/wsxiaoys/e0286dc6bb624ff5fdf49e7f4c528ba3
How reasoning prefill works
The core trick is called "reasoning prefill." You inject the target model's chain-of-thought prefix at the top of the prompt, and Qwen continues along the same line of thinking.
The prompt template from the Gist looks roughly like this:
You are a helpful assistant.
<reasoning_trace>
The user is asking about X. Let me think step by step...
First, I need to consider...
Then, I should evaluate...
Finally, the answer is...
</reasoning_trace>
Now answer the following question:
[user's actual question]The prefill section contains three parts:
- Role instruction. Tells the model who it is.
- Target model's reasoning prefix. Pulled directly from GPT-5.5 Pro's output, including its chain-of-thought markers like "let me think...," "first...then...finally..."
- Continuation cue. Signals the model to keep writing in that vein.
Why does this work? LLMs are highly sensitive to preceding tokens during generation. Once the context contains traces of the target model's reasoning, Qwen starts imitating that style instead of defaulting to its own format.
wsxiaoys found that after receiving GPT-5.5 Pro's reasoning prefix, Qwen 3.8 produces output with formatting, word choice, and even paragraph structure that closely match GPT-5.5 Pro.
How this differs from existing approaches
Distillation, few-shot, prompt templating — each has its limits:
- Distillation: Requires compute, data, and training. Fine-tuning Qwen on GPT-5.5 Pro's outputs gives the best results but costs the most.
- Few-shot: Stuffing a few examples into the prompt teaches format but only surface patterns, not deeper reasoning style.
- Prompt templating: Hardcoding format instructions like "answer in the following format." Weakest option, and models often ignore it.
Reasoning prefill sits between few-shot and distillation. No training required, yet it captures deeper reasoning patterns. Essentially, it "poisons" the model's context into thinking the target style is its own.
The bigger picture: open-source style is a copyable asset
This is where it gets interesting.
Big companies spend serious money on brand voice. OpenAI's crispness, Anthropic's rigor, Google's verbosity. These style differences used to be part of the moat.
Prefill proves that style can migrate at zero cost.
For closed-source vendors, the implication is clear: a brand voice that cost millions to tune can be replicated with a single prompt. The moat shifts from style to capability — whether you can solve harder problems, not whether your output sounds a certain way.
For the open-source community, this further unlocks the potential of models like Qwen, Llama, and Mistral. No need to wait for official fine-tunes; a prompt gets you the style you want.
Tested limits: where it works and where it breaks
Based on the HN discussion and Gist comments, here's where the boundaries sit:
Works well for:
- Code completion (consistent reasoning style leads to consistent code style)
- Structured reports (format imitation is strong)
- Long-document summarization (chain-of-thought transfer holds up)
Breaks down for:
- Prefill exceeding the context window — it simply won't fit and fails outright
- Reasoning paths that depend on tool calls. If GPT-5.5 Pro's trace includes function calls, Qwen can't necessarily reproduce them
- Multi-turn style drift. By turn five, Qwen gradually reverts to its default style
There's also a hidden issue: if GPT-5.5 Pro's reasoning contains an error, Qwen carries the mistake forward rather than correcting it.
Editorial take
This raises a question about what "style moat" even means going forward.
The gap between open and closed models used to be framed as capability. Closed models were stronger. Now it looks like style is getting flattened too. Competition narrows to a single dimension: can you solve the problem?
My read: in the second half of 2026, more vendors will productize this trick. Not as something the open-source community does quietly, but as a SaaS feature — pick a target model, and the tool generates a prefill prompt that makes your model sound like it. Tools like Cursor, v0, and Devin already do model routing; adding a "style router" isn't a stretch.
What's different here is the goal. Traditional prompt engineering aimed at helping the model understand the task. This aims at getting the model to pretend to be a different model. The objective shifts from understanding to imitation, and the attack surface expands from output content to output style.
Open questions: does the Qwen team know about this? Will they add defenses in the next version — like detecting prefill injection and resetting context — or lean into it as an official feature? No public statement from Qwen so far.
Another gap: wsxiaoys only tested Qwen 3.8. How does this work on Llama 4 or Mistral Large? No comparison data in the Gist.
Regardless, one thing is clear: open-source models are more malleable than most people assume. No need to wait for the next official release — a prompt unlocks new capabilities.