Matt Pocock's "Dictionary of AI Coding" Gives Your LLM a Shared Vocabulary
Watch on TikTok
Matt Pocock released an open-source glossary called dictionary-of-ai-coding that defines every major concept in AI-assisted development using plain English. The project is hosted on GitHub and organized into seven sections covering models, context windows, tools, failure modes, handoffs, memory, and patterns of work. The twist is that it is designed to be injected directly into your LLM's context window so your agent uses consistent, industry-standard terminology instead of making up its own.
What the Dictionary Actually Contains
The repository at github.com/mattpocock/dictionary-of-ai-coding covers the full surface area of working with AI coding agents. Section 1 ("The Model") defines foundational terms like model, parameters, training, and inference with precise, opinionated definitions. Each entry follows the same structure: a one-paragraph definition, usage examples written as realistic conversations between a developer and a teammate, and an "Avoid" note that flags common misuses of the term.
Later sections get more practical. Section 4 (Failure Modes) includes terms like "hallucination" and explains why it happens. Section 5 (Handoffs) defines concepts like "human review," "vibe coding," and "design concept," drawing clear lines between them. For example, the dictionary explicitly warns against using "vibe coding" as a synonym for "low-quality AI code," noting that the term describes the review stance (accepting code without reading it), not the code quality itself.
Why It Is Formatted as an Agent-Native Reference
This is where the project gets interesting. The dictionary is not just documentation for humans. It is structured as a reference module you inject into your LLM's context. The repo includes a CLAUDE.md file, which means tools like Claude Code will automatically pick up these definitions when working inside the repository.
The practical benefit: when you and your agent both use the same definitions for "context window," "tracer bullet," or "agentic harness," you spend less time correcting misunderstandings and more time building. Pocock's argument is that most confusion around AI coding is manufactured by jargon, and giving your agent strict definitions eliminates an entire class of communication failures.
The Sections at a Glance
The dictionary is organized into seven sections, each covering a distinct area of AI-assisted development:
- The Model covers what a model is, how parameters and training work, and what inference means in practice.
- Sessions, Context Windows & Turns defines the boundaries of how LLMs process and retain information.
- Tools & Environment explains the infrastructure that wraps around a model to make it useful.
- Failure Modes catalogs the ways AI coding goes wrong, with precise definitions for each.
- Handoffs covers the human-AI boundary, including human review, automated review, and vibe coding.
- Memory and Steering addresses how agents retain and apply context across sessions.
- Patterns of Work describes repeatable workflows for AI-assisted development, like grilling and design concepts.
Key Takeaways
- Matt Pocock's dictionary-of-ai-coding is an open-source glossary that standardizes AI coding terminology in plain English.
- Each entry includes definitions, usage examples, and "Avoid" notes that flag common misuses of terms.
- The dictionary is designed to be injected into your LLM's context, so your agent uses the same vocabulary you do.
- Seven sections cover the full lifecycle of AI-assisted development, from model basics to handoff patterns.
- The project had 243 stars and 54 forks at the time of the video, with active development (commits within minutes of recording).
Resources
Published June 11, 2026. Writeup generated from a favorited TikTok.