Shared Language Glossary
A project glossary file gives the agent and the human one ubiquitous language, which cuts verbosity, keeps naming consistent, and doubles as a source of leading words.
Evidence: moderate
Matt Pocock borrows the ubiquitous language from Domain-Driven Design. An agent dropped into a project has to infer its jargon, so it uses twenty words where one would do. A glossary file (his GLOSSARY.md, formerly CONTEXT.md) fixes this: each term gets a definition, the synonyms to avoid, and relationships between terms, and flagged ambiguities record how a confusion was resolved. His own example compresses a sentence about a lesson being given a place on the filesystem into "the materialization cascade".
He claims three payoffs beyond brevity: variables, functions and files get named consistently, the codebase becomes easier for the agent to navigate, and the agent spends fewer tokens thinking because it has a more concise language. These are practitioner claims shown in his repositories rather than measured.
The glossary is also a pool of leading words. When the same term lives in prompts, docs, skills and code, a context pointer using it fires more reliably. wait-what, a three-line corrective for an over-verbose message, works by reusing that vocabulary.
The repository distinguishes passive use (reading the glossary, a one-line pointer in other skills) from active domain modelling (challenging terms, edge-case scenarios, updating the glossary and decision records inline), which is its own skill.
Boundary: a glossary that restates code identifiers adds nothing, because it becomes a cache of what the code already says; its value is terms and decisions the code cannot express.
Sources
- https://github.com/mattpocock/skills (read 2 Oct 2026, v1.2.3)
- https://github.com/mattpocock/skills/blob/main/CHANGELOG.md (read 2 Oct 2026)