Reference · AI Hero
Matt Pocock's vocabulary, in his wording. Once a term is here, every lesson uses it this way.
A term marked speaker background is not on any slide. Keep it for a question from the room. Every other term is used in the deck.
| Term | Meaning | Source |
|---|---|---|
| AI engineer | “A software developer who builds applications powered by AI.” Builds applications that use AI models via APIs. | what-is-an-ai-engineer |
| ML engineer | Builds the AI APIs themselves. Model training and optimisation. The other side of the API boundary. | what-is-an-ai-engineer |
| Agent | “Doesn't use predetermined steps. It calls an LLM and gives it a bunch of tools.” The LLM decides when to stop. | what-is-an-agent |
| Workflow | Pre-defined code paths. “The code itself decides when to stop the program, when to call the next LLM.” | what-is-an-agent |
| Plan mode | “A restricted mode”. The agent reads files and runs shell commands to explore. It does not edit your source until you approve the plan. Edits are blocked, not commands. | plan-mode-introduction, code.claude.com/docs/en/permission-modes |
| Context window | Everything the agent can see this turn. “An agent can only work with what’s in its context window.” | plan-mode-introduction |
| Session | “Whatever isn’t in the session doesn’t exist for the agent.” Everything the agent knows about the conversation, held in its context window. | ai-coding-dictionary/session |
| Token | The unit the context window is measured in. “a thousand tokens is roughly 750 words”. | ai-coding-dictionary/token |
| Attention budget | “LLMs have an ‘attention budget’ that they draw on when parsing large volumes of context. Every new token introduced depletes this budget by some amount.” The model does not forget — the signal gets lost in the noise. | anthropic.com/engineering/effective-context-engineering-for-ai-agents, ai-coding-dictionary/attention-budget |
| Context rot speaker background |
“As the number of tokens in the context window increases, the model’s ability to accurately recall information from that context decreases.” A gradient, not a cliff. | anthropic.com/engineering/effective-context-engineering-for-ai-agents |
| Smart zone / dumb zone | “Early in a session the agent is in a “smart zone” — sharp, focused, recall is good. As the session grows it drifts into a “dumb zone”: sloppier, forgetful, more mistakes”. The zones do not follow the context window limit. | ai-coding-dictionary/smart-zone |
| Compaction speaker background |
“A handoff done in-memory: the previous session’s history is summarised, and the summary seeds a fresh session. Lossy by design”. “Compaction is lossy, and autocompact is lossy at a moment you didn’t choose.” | ai-coding-dictionary/compaction, ai-coding-dictionary/autocompact |
| Subagent | “Each subagent starts with a fresh, isolated context window” and “returns only a condensed, distilled summary of its work.” Exploring a codebase this way does not spend the main context. | code.claude.com/docs/en/sub-agents, anthropic.com/engineering/effective-context-engineering-for-ai-agents |
| CLAUDE.md | “CLAUDE.md is a special file that Claude reads at the start of every conversation”. Persistent context it can’t infer from code alone — but “target under 200 lines per CLAUDE.md file.” | code.claude.com/docs/en/best-practices, code.claude.com/docs/en/memory |
| /init | “Run /init to generate a starting CLAUDE.md automatically”. The place to start the file, not to finish it. |
code.claude.com/docs/en/memory |
| /clear | “Clearing is the cure for a polluted context.” Resets the window entirely — the move once you have corrected the same issue more than twice. | ai-coding-dictionary/clearing |
| Slash command | A command you type into the agent, starting with /. /clear and /init are two. A skill is called the same way: “Custom commands have been merged into skills.” |
code.claude.com/docs/en/skills |
| Permission prompt | The question the agent asks before it acts, as shown on slide 4. You approve the action, or you refuse it. “Claude Code only has the permissions you grant it. You’re responsible for reviewing proposed code and commands for safety before approval.” | code.claude.com/docs/en/security, code.claude.com/docs/en/permissions |
| Faithfulness hallucination | “The knowledge is present but losing the competition for attention: the fix is removing context.” Paste in more docs, and the drift gets worse — the beginner’s trap. | ai-coding-dictionary/hallucination |
| Test | Code that runs your code and checks the answer, so the result is one word: pass, or fail. Slide 5 shows one, with the code it checks. It is the feedback loop the deck teaches first, because the agent can run it and read the answer without asking you. “Basic unit tests covering core functionality help keep the AI on track.” | essential-ai-coding-feedback-loops |
| Feedback loop | Any automatic pass-or-fail signal “so the AI can verify its own work”. A typecheck, a test run, a blocked commit. Slides 6 to 8 show one closing: the check fails, the agent reads the failure, fixes it, and a pass ends the loop. Slide 21 lists four to install. | essential-ai-coding-feedback-loops |
| Skill | “A teachable capability bundled as a unit — instructions and resources for doing one task well, kept in the environment until a context pointer pulls it into the context window for the task at hand.” “Only the name and description sit in context by default.” | ai-coding-dictionary/skill |
| Context pointer speaker background |
“A pointer is one line in the context window. The document behind it might be thousands of tokens, but those tokens cost nothing until the agent actually follows the pointer.” | ai-coding-dictionary/context-pointer |
| SKILL.md | “Skills extend what Claude can do. Create a SKILL.md file with instructions, and Claude adds it to its toolkit.” “Keep SKILL.md under 500 lines.” | code.claude.com/docs/en/skills |
| /ask-matt | “Find out which skill to use for the situation you are in.” The router over all 25 skills. | aihero.dev/skills |
| /grill-me | “Align on an idea before committing to it.” It talks. It writes nothing down. Slide 22 shows it in full — three sentences. | aihero.dev/skills |
| /grill-with-docs | “Get interviewed about a plan, and record the decisions.” The same interview, but it leaves files on disk. Stage one of the main flow. | aihero.dev/skills |
| /to-spec | “Turn an agreed conversation into a written spec.” Stage two. The spec is what survives the session. | aihero.dev/skills |
| /to-tickets | “Split a spec into small tickets an agent can build.” Stage three. Each ticket is a fresh session that reads the spec. | aihero.dev/skills |
| /implement | “Build a finished spec into code, test-first.” Stage four. /tdd is the engine inside it. |
aihero.dev/skills |
| /tdd | “The rules of the red-green-refactor loop.” Not a stage of its own. It runs inside /implement. |
aihero.dev/skills |
| /code-review | “Review a diff against your standards and against the spec.” Stage five. It does not replace you reading the diff. | aihero.dev/skills |
| Spec | “The spec exists because context windows end.” | skills-to-spec |
| Harness | “The agent loop — model proposes, harness executes, repeat — is run by the harness.” | ai-coding-dictionary/harness |
| Eval | “Evals are the AI engineer's unit tests. They are how you wrangle predictability from a probabilistic system.” Produces a score, not a pass or fail. | what-are-evals |
| Deterministic eval | Returns “a pass/fail result”. Seeks “to extract determinism from a probabilistic system”. | three-types-of-evals |
| LLM-as-a-judge | An LLM scoring another LLM's output, including factuality against a ground truth. A first screen, not a final test. | three-types-of-evals |
| Human eval | “Some evals can only be usefully evaluated by humans.” Long-form generation. “Human oversight is needed for any type of LLM app.” | three-types-of-evals |
| Staircase Of Complexity Hell | 18 techniques ordered by cost. Simple at the top. “Start at the top of the staircase, and work your way down only when you've exhausted all the simpler options.” | how-to-improve-your-llm-powered-app |
| Deep module | “Lots of implementation controlled by a simple interface.” | how-to-make-codebases-ai-agents-love |
| Grey box architecture | “You own the interface. AI owns the implementation. Tests keep it honest.” | how-to-make-codebases-ai-agents-love |
| Progressive disclosure speaker background |
“Skills are the pattern built into the harness: a short description loaded every session, the full instructions only when triggered.” | ai-coding-dictionary/progressive-disclosure |
| Hook | Two different things share the name. A git hook (for example a pre-commit hook) fires on a git action: “If any step fails, the commit is blocked and the AI gets an error message.” A Claude Code hook is a “user-defined shell command” that Claude Code runs at points in its own lifecycle, “which gives you deterministic control”: PostToolUse fires “After a tool call succeeds”; Stop fires “When Claude finishes responding”. The loop on slide 21 is the git hook; the Claude Code hook is no longer its own slide — it is the spoken aside next to slide 21. |
essential-ai-coding-feedback-loops (git hook), code.claude.com/docs/en/hooks-guide (Claude Code hook) |
| Instruction budget speaker background |
Quoting Kyle at Humanlayer: “Frontier thinking LLMs can follow ~150-200 instructions with reasonable consistency.” Not Matt's own measurement. | a-complete-guide-to-agents-md |
| Vertical slice speaker background |
A ticket that cuts through every integration layer, rather than a horizontal architectural slice. What /to-tickets produces. |
5-agent-skills-i-use-every-day |
| Tool | A way “of getting LLMs to produce certain types of messages which you can then intercept and execute on your machine”. Defined by name, description and parameters. | what-are-tools |
| Structured output | Passing a schema “directly to the generateObject function from the AI SDK” and reading back a typed, validated object. |
structured-outputs-with-vercel-ai-sdk |
| Evalite | “It's based on Vitest, and it's called Evalite.” A local-first, TypeScript-native tool for testing LLM-powered apps. No API key needed. | evalite-an-early-preview |
| TTFT / TPOT | Time to first token, and time per output token. The two latency numbers when choosing a model. | how-to-choose-an-llm |