Semantic Anchors

The site https://llm-coding.github.io/Semantic-Anchors/ provides an extensive list of what they call semantic anchors.

A semantic anchor is a word or short phrase that has been established to refer to a wider concept. Instead of bringing the exhaustive description to the prompt in a LLM conversation, you can just use the short term.

E.g. when you what to have your development happen in a test driven manner, you would normally describe the process of writing a failing test first, then writing the implementation until all tests all green and then do the refactoring.

The semantic anchor approach is to just say TDD, Chicago school.

What that means is expected to be already contained in the training set of the large language model. Because most billing is done on a per token basis, this leads to less token usage. And less tokens also means less distraction for the model.

The site Semantische Anker verkürzen den Kontext für das agentische Coden also describes the feature in German.