When writing prompts and skills, there is a highly counterintuitive phenomenon: many people like to write rules using "a small label plus a colon plus a detailed description," believing this provides a clear outline and structure. In reality, it instead causes a sharp drop in model instruction compliance.
This style is very commonly used, and due to the influence of training corpora, almost all models use it by default, for example:
- Exception handling: When an error occurs, throw it upward directly; do not catch it at the current layer
- Core principles: ...
- Input validation: ...
It looks typographically clean, with a clear outline—like a very good document. In human reading habits, this is using leading words to highlight key points. However, during actual execution, you will find that the AI rarely follows the content after them; most of its attention is on the leading words before them.
For example, general terms like "exception handling" appear tens of thousands of times in the corpus, and what they are bound to are all the most mediocre generic code patterns. Once the model's front-loaded attention is hijacked by this word, it instantly activates the default template, and the specific counter-constraints behind it simply cannot outcompete the prior weights in front.
In addition, colon structures in the training set largely correspond to official documents, encyclopedias, and reference documents, so the model's inertia treats them as passive knowledge rather than an enforced protocol.
The truly effective approach is to eliminate all colons and labels and change them into flat action instructions.
For example, directly write "When an error occurs, it must be thrown upward; catching it at the current layer is strictly prohibited," with the condition first, the action following, using declarative sentences.
Writing rules does not need an outline feel; the more direct and crisp the instructions, the more precise the execution.