W34 - Some AI Practice Shares

The most intuitive use is multi-model collaboration: pose questions with Gemini-2.5-pro and have Claude write the response, effectively having Claude work for Gemini. This pairing lets each model play to its strengths, producing higher efficiency and better output quality in the question-and-answer flow.

Input quality directly determines output quality. Structured expressions are often far more efficient than casual natural-language descriptions, making it easier for the model to grasp key points. Role prompts are not decorative: in testing and coding scenarios, role definitions produce noticeable differences. In coding tasks specifically, model performance depends more on the completeness of context; compared with role settings, providing sufficient contextual information in the input has a greater impact on results.

To further understand prompting, take it seriously as an engineering discipline. Compared with resource-intensive methods like fine-tuning, prompt engineering can deliver performance leaps in much less time. Looking at leaked system prompts from various vendors, many run to thousands of lines — essentially engineering code written in natural language. They contain processes, constraints, and priorities, functioning exactly like large configuration files.

Different models show clear preferences for prompt formats. OpenAI's GPT series is more stable and logically consistent with JSON-structured input, while Anthropic's Claude series performs more robustly with XML or tag-style formats. Feeding the same prompt to GPT and Claude prompt optimizers and comparing the resulting styles gives a direct sense of each vendor's model preferences.

Regarding account bans, I think Claude enforces the strictest user policies. An account I registered last year was banned in April, with a remaining balance and no explanatory email. I did discover a useful trick: register using Apple iCloud's private relay email. Third parties then only see Apple's relay service, making identity tracking much harder.

On the Claude Code experience: its custom slash commands and hooks are powerful, highly extensible, and easy to migrate. Claude Code (CC) can integrate with mainstream IDE terminals for a smooth experience. My current go-to setup is Windsurf paired with Claude Code, which preserves the feel of the IDE while fully leveraging CC's extension capabilities.