domain-modeling

Most misbuilds start as a language gap: the agent is dropped into a project and left to infer the jargon, so it uses twenty words where the domain has one. A shared language closes the gap. When code, conversation, and the model all draw from the same vocabulary, names line up, navigation gets cheaper, and the model spends fewer tokens reasoning because it has a tighter language to reason in.

Method

1. Harvest the terms. From the request, the codebase, and the user's own words, list the nouns and verbs that carry domain meaning - the concepts a newcomer would have to ask about. Prefer the user's word over a synonym you like better.
2. Pin each one. Write a one-line definition in the project's own language, not a dictionary definition. If two terms blur together, force the distinction or collapse them - ambiguity here becomes inconsistent names in code.
3. Draw the boundaries. Where the same word means different things in different parts of the system, that is a boundary. Name each context and note which terms belong to it. A term that means two things is two terms.
4. Record the hard calls. When a modeling choice was contested or will be questioned later, write a short decision record: context, choice, alternatives rejected, why.

Outputs

• CONTEXT.md - the shared-language glossary. One term per line: term - what it means in this project. Grouped by bounded context when there is more than one. Point every future session at this file. On re-run, add new terms and update definitions that changed; do not rewrite the file wholesale.
• Bounded-context sketch - the contexts and which terms live in each, short enough to read in fifteen seconds.
• Decision records in docs/decisions/NNNN-slug.md for the contested modeling calls only. Read the directory first and number from the highest existing record so two records never collide. Skip the obvious ones.

Guardrails