Semantic anchors shorten the context for agentic coding
Clean Architecture, Code Smells or Wolf Schneider – a directory on GitHub collects such semantic anchors for short prompt context.
The Semantic Anchors directory on GitHub lists semantic anchors, i.e., keywords with which prompters can provide LLMs with complex concepts in short terms. The page sorts the anchors into categories such as “Design Principles & Patterns,” “Development Workflow,” or “Documentation Practice.”
For each of these keywords, the explanation of the underlying concept can be found in the directory. For easier searching, roles can be filtered to show only relevant anchors, for example “DevOps Engineer” or “Software Developer/Engineer”.
Videos by heise
The operator Ralf D. MĂĽller regularly adds new anchors and accepts suggestions, for example, via issue. The submitter should first familiarize themselves with the quality criteria.
LLMs know the concepts
The idea behind semantic anchors is that LLMs already know extensive concepts, and developers therefore do not need to explain them in detail in the prompt. For example, “Code Smells by Fowler” is enough of a hint for the LLM to clean up many weaknesses in the code during a review. For a properly formulated text, keywords such as “Good German according to Wolf Schneider” or “Plain English according to Strunk & White” are useful.
The page also introduces other auxiliary structures for LLM use, such as semantic contracts, which define fundamental terms and guidelines and can be integrated into AGENTS.md. Furthermore, MĂĽller describes in detail an agentic, spec-driven workflow based on anchors.
(who)
