Stop explaining TDD to your agent — just name-drop it

What it does

Semantic Anchors is a browsable catalog of 230+ established terms and methodologies — things like TDD London School, arc42, tracer bullet — chosen because they reliably activate rich, specific knowledge inside an LLM. The idea: instead of spelling out “write tests first, use mocks, design outside-in” in your agent instructions, you say “TDD, London School” and let the model’s training data do the rest. The catalog spans 12 categories, from software architecture to requirements engineering, and can be filtered by professional role.

The interesting bit

It’s essentially a prompt-engineering dictionary with quality control: each candidate anchor is run through an “Anchor Prior Test” against criteria like precision, consistency, and attributability (can it be traced to key proponents or publications?). There’s also a pragmatic admission that smaller local models may not know all the anchors — the project supports injecting explicit definitions into agent instruction files for Claude Code, Codex, Gemini CLI, Cursor, Copilot, and Windsurf.

Key highlights

• 230+ anchors across 12 MECE categories, filterable by 15 professional roles
• Interactive website with search, card-grid visualization, and English/German switching
• Contribution pipeline where you propose a term, maintainers test it, and Copilot drafts the entry
• ADR-008 documents a “utility gate” for personality/assessment frameworks — they’re picky about what gets in
• The repo itself is a Vite static site; the content is the product

Caveats

• The core claim — that a named term activates “deep, consistent” knowledge — is asserted rather than benchmarked in the README; results will depend heavily on the model you use
• It’s a catalog, not a tool: you still have to wire the anchors into your own prompts or agent configs

Verdict

Worth a browse if you write agent instructions or system prompts and find yourself re-explaining the same methodologies. If you were hoping for a library or API, this is vocabulary, not code.