DoneHub

Same Model, Different Experience

Here’s an odd thing: use GPT-4 with OpenAI’s Codex CLI and you get decent results. Use GPT-4 with Anthropic’s Claude Code and you get better results. You’d expect a model plus its own vendor’s agent to be the best combo — but it actually underperforms paired with a competitor’s agent.

The common assumption is that Claude Code just has a better prompt. Prompts matter, but they’re not the deciding factor. Claude Code’s source reveals something more fundamental: it’s not a prompt wrapper. It’s a full agent runtime. This framework understands the model’s capability boundaries, tool execution risks, context compression strategies, and intent flow. The model is the brain, and Claude Code is the body. The analogy is rough, but it captures the core relationship.

On April 22, 2026, Anthropic published an official blog post about MCP’s positioning and evolution. The context was anything but calm — over the preceding months, the community had raised serious concerns about MCP’s cost, context consumption, and schema design. Perplexity’s CTO even publicly stated they were moving away from MCP internally. Anthropic’s response didn’t rebut each criticism point by point. Instead, it offered a more complete framework for understanding how agents connect to external systems, along with two concrete technical optimizations that directly addressed the community’s core concerns.

Read more »

I was using Claude Code to refactor a user module. The first 30 rounds went smoothly — it remembered clearly that the database was PostgreSQL, the API style was RESTful, and the Redis cluster was under maintenance that day so sessions should fall back to the DB temporarily. Then around round 50, it suddenly asked, “Can’t find the MySQL MCP info, could you provide it?” The MCP settings were right there in the file — it had known about them earlier. Once the context window gets overloaded, early critical constraints get diluted in the model’s attention. This article breaks down exactly what happens inside Claude Code when you hit /compact.

You want AI to help debug a web page, but it can only offer suggestions — it can’t actually do anything. You need to automate testing a page’s functionality, but writing scripts feels like overkill. You’re analyzing performance bottlenecks and the DevTools panel is a wall of metrics you can’t parse. The common thread across these scenarios: there’s no bridge between AI and the browser. Chrome DevTools MCP builds that bridge. It lets AI assistants like Claude Code, Cursor, and Copilot directly operate the browser — working like an actual developer would.

In April 2026, China’s AI compute supply was tight enough to feel like rationing. Signing up for a Coding Plan meant a long waitlist just to get a slot. The reason is straightforward: high-end GPUs are banned from export, and domestic alternatives can’t keep up. This article breaks down the real constraints facing China’s GPU self-development effort across three dimensions: fabrication process, software ecosystem, and talent pipeline.

Read more »

It Starts with a Compute Shortage

In 2026, scrambling for compute has become the norm in the AI industry. China’s compute crunch traces back to one dependency: AI needs GPUs, and for high-end GPUs, NVIDIA is virtually the only option. Many people know NVIDIA sells graphics cards, but fewer are familiar with CUDA. CUDA is NVIDIA’s most critical competitive moat — one that took twenty full years to build. Understanding how CUDA got here is the key to understanding today’s compute market.

Read more »

Background

Ask an LLM “Who was the first emperor of the Ming Dynasty?” and it will answer “Zhu Yuanzhang.” But ask “What was the name of Zhu Yuanzhang’s third wife?” and it might answer “Empress Ma.” Empress Ma was Zhu Yuanzhang’s principal wife, not his third wife — the model has generated an answer that sounds plausible but is factually wrong. This is hallucination: the model speaks with confidence while getting the facts wrong. In a job interview, saying “because it’s a probabilistic model” won’t cut it. A probabilistic model doesn’t inevitably fabricate content — you need to unpack which specific stages in the pipeline cause the problem.

Read more »

A video circulating in tech circles the other day listed four core skills for AI Agent engineers: programming fundamentals, Prompt Engineering, systems thinking, and domain knowledge. The comments section was full of people bookmarking it saying “so true.” I don’t dispute that framework, but after shipping several Agent projects to production, I’ve found the real bottleneck skills are ones that video never mentioned.

When describing how Function Calling works, many people say the model “reads” the user’s intent, “determines” it needs to call a tool, then “decides” to generate a call instruction. This explanation is popular, but words like “reads,” “determines,” and “decides” are misleading — they imply the model possesses some understanding ability, as if it’s actually thinking. An LLM’s nature has never changed: it’s Next Token Prediction. Function Calling doesn’t introduce any mysterious reasoning mechanism. It just teaches the model to switch output formats under specific conditions — from natural language to structured JSON text.

Read more »

Building a knowledge base Q&A system in an enterprise usually comes with a choice: use standard vector-retrieval RAG, or use Microsoft’s GraphRAG? The two differ significantly in technical approach and suitable scenarios. Picking the wrong one means either insufficient retrieval accuracy or excessive engineering complexity. This article analyzes the capability boundaries and best-use timing of both approaches from a practical scenario perspective.