# How does an AI tool learn our internal acronyms and jargon instead of using the general meaning?


URL: https://getunblocked.com/answers/teach-ai-internal-acronyms-and-jargon/
Published: 2026-11-10T15:30:00Z
Updated: 2026-11-08T16:32:19Z
Author: Dennis Pilarinos
Categories: Context Engine, Engineering Insights

**Short answer:** It doesn't learn them on its own. The model only knows the public meaning, so you have to put your meaning in front of it. Write a short glossary into the rules file your tools load, such as CLAUDE.md, Copilot's instructions file or Cursor rules. For the long tail, connect a tool that finds how your team actually uses each term in PRs, tickets and chat, and cites it.

---
An AI tool does not learn your acronyms by being used. Each session starts from public training data, so "PSR" or "the Atlas cutover" gets its most common public meaning unless you put your meaning in front of the model every time it answers. **Internal jargon** is the set of acronyms, code names and loaded words that mean something specific inside one company and something else, or nothing, outside it. A short glossary in the rules file your coding tools already load covers the core terms. Retrieval over your PRs, tickets, docs and chat covers the rest, as long as it cites where each meaning came from.

## Why the model picks the public meaning

A language model has no record of your company, your services or the names your teams gave to projects over the years. When it meets an unfamiliar acronym, such as a three-letter service name, it reaches for the meaning it saw most often in training and states that guess with the same confidence as a fact.

A March 2026 study of clinical notes shows how wide that gap can be. Small general-purpose models running on-device spotted acronyms with about 0.988 accuracy, but "their expansion capabilities plummet (\~0.655)" when asked what the acronyms stood for. Routing the acronyms to biomedical models lifted expansion to about 0.81 ([Aithal et al., arXiv](https://arxiv.org/abs/2603.23678), 2026). Medicine is not software, and your numbers will differ, but the pattern carries over: the model recognizes an acronym long before it knows what your team means by it.

The tools say this plainly. Cursor's docs open with "Large language models don't retain memory between completions" ([Cursor docs](https://cursor.com/docs/context/rules)), and Claude Code's docs say "Each Claude Code session begins with a fresh context window" ([Claude Code docs](https://code.claude.com/docs/en/memory)). Whatever you explained to the agent yesterday is gone today, unless it is written down somewhere the agent reads at the start of every session.

## Start with a glossary in the rules file

**A rules file** is a plain Markdown file in the repository that a coding tool adds to the start of every session, such as `CLAUDE.md` for Claude Code. It is the cheapest place for a glossary. Keep it to the terms that come up in most tasks, with one line each giving the term, what it means here, and where it lives in the code.

- Claude Code reads `CLAUDE.md`, and these files are "loaded at the start of every conversation" ([Claude Code docs](https://code.claude.com/docs/en/memory)).
- GitHub Copilot reads `.github/copilot-instructions.md`, which GitHub describes as "repository-specific guidance and preferences" ([GitHub docs](https://docs.github.com/en/copilot/how-tos/configure-custom-instructions/add-repository-instructions)).
- Cursor reads project rules from `.cursor/rules`, and a rule set to Always Apply is included in "every chat session" ([Cursor docs](https://cursor.com/docs/context/rules)).

For example, a made-up entry might read: "PSR: payment settlement run, the nightly batch in `billing/settlement/`, owned by the billing team, not the PHP standards." That single line keeps the agent from offering PHP style advice inside a Go service. If your teams use more than one tool, keep one glossary and point each file at it; our answer on [keeping rules files in sync across repos](https://getunblocked.com/answers/sync-rules-files-across-repos/) covers how.

## Where the glossary stops working

A rules file is a good home for 30 terms and a poor home for 400. Claude Code suggests you "target under 200 lines per CLAUDE.md file," because "longer files consume more context and reduce adherence" ([Claude Code docs](https://code.claude.com/docs/en/memory)), and Cursor says to "keep rules under 500 lines" ([Cursor docs](https://cursor.com/docs/context/rules)). Claude Code also treats these files "as context, not enforced configuration," so a glossary nudges the model rather than binding it. If an agent keeps ignoring an entry, our answer on [why agents ignore CLAUDE.md rules](https://getunblocked.com/answers/agents-ignore-claude-md-rules-how-to-make-context-stick/) goes through the fixes.

Jargon also changes faster than anyone edits a glossary. A code name gets reused for a new project, a renamed service keeps its old name in Slack for a year, and the same acronym means one thing to the payments team and another to the data team. A hand-written list goes stale the week someone forgets to update it, and nothing warns the agent.

|  | Glossary in a rules file | Cited retrieval |
| --- | --- | --- |
| Good for | 20 to 30 stable, common terms. | The long tail, renamed and reused terms. |
| Size limit | Under 200 lines (Claude Code), under 500 (Cursor). | Whatever your sources hold. |
| Stays current | Only when someone edits it. | Reads the latest PRs, tickets and chat. |
| How you check it | Read the file. | Follow the cited link. |


## Let the tool find the meaning where your team uses it

Most internal terms are already defined somewhere. The trouble is that it is never one place: the PR that introduced the service, the ticket that named the project, the incident thread where someone spelled it out, and the code itself. **Cited retrieval** is when a tool searches those sources at question time and returns the meaning with links to where it found it. That is how a teammate would answer "what does PSR mean here?"

That works when the tool searches code, PRs, tickets, docs and chat together, because the definition is rarely where the question is. It also has to prefer recent, authoritative sources, so last month's rename beats a two-year-old wiki page ([how the Unblocked context engine ranks sources](https://getunblocked.com/blog/inside-the-unblocked-context-engine/) explains one approach). And it has to cite, so an engineer can check the meaning in seconds. We wrote about the wider problem in [what your coding agent can't see](https://getunblocked.com/blog/what-your-coding-agent-cant-see/).

That is the job Unblocked does. Its `context_research` tool runs "multi-source research across code, PRs, docs, issues, and messaging" and "returns a synthesized answer with citations" ([Unblocked docs](https://docs.getunblocked.com/unblocked-mcp/mcp-overview.md)). Answers reach Claude Code, Cursor, Copilot, VS Code and JetBrains over MCP, and Slack through the Unblocked app. Each person sees only sources they can already access ([Data Shield](https://docs.getunblocked.com/team-settings/data-shield)). For how terms and owners link up across systems, [knowledge graphs for engineering teams](https://getunblocked.com/blog/knowledge-graph-for-engineering-teams/) compares the options.

## What to do this week

- List the 20 to 30 terms new hires ask about most. Put them in one glossary file with a meaning and a code path each, and link it from `CLAUDE.md`, `.github/copilot-instructions.md` and your Cursor rules.
- Pick ten terms that are not on the list, ask your agent what each one means in your codebase, and check the answers.
- For every wrong answer, note where the right meaning actually lives, whether a PR, a ticket or a thread.

If the glossary clears all ten, you are done. If the agent keeps guessing, the missing meanings live in systems it cannot read yet, and that is where retrieval with citations earns its place.

## Read more

- [What your coding agent can't see](https://getunblocked.com/blog/what-your-coding-agent-cant-see/)
- [Knowledge Graph Tools for Engineering Teams: What to Use in 2026](https://getunblocked.com/blog/knowledge-graph-for-engineering-teams/)

## Related answers

- [Is an LLM-maintained wiki in git a better fix than a context tool for engineering knowledge?](https://getunblocked.com/answers/llm-maintained-wiki-vs-context-tool/)
- [How do we drive adoption of an AI coding tool from a third of engineering to everyone?](https://getunblocked.com/answers/drive-ai-coding-tool-adoption-across-engineering/)
- [How do I make sure Claude actually calls the context MCP without every engineer editing their CLAUDE.md?](https://getunblocked.com/answers/make-claude-call-context-mcp-without-editing-claude-md/)
- [Atlassian Rovo MCP versus a dedicated context tool for Confluence, Jira and code?](https://getunblocked.com/answers/atlassian-rovo-mcp-vs-context-tool/)