A-Z Glossary
All the important terms in one place, to scan in 5 minutes before the interview.
| Term | Meaning |
|---|---|
| Adaptive thinking | Claude decides for itself how much to think; thinking: {type: "adaptive"} |
| Agent | An LLM that uses tools in a loop and decides the next step itself |
| Agent SDK | Claude Code's harness as a Python/TS library |
| Artifact | Code/page/document/app built in a separate window in claude.ai |
| ASL | AI Safety Level: the safety tiers of the RSP |
| Batch API | Async requests, 50% cheaper |
| Cache breakpoint | A cache_control marker; the prefix up to here is cached |
| Citations | Exact quotes/pages from the document along with the answer |
| CLAUDE.md | Claude Code's project memory file |
| Compaction | A server-side summary of the old conversation |
| Computer use | Looking at screenshots and driving the mouse/keyboard |
| Constitutional AI | Self-critique based on principles + RLAIF training |
| Context editing | Clearing old tool results/thinking |
| Context window | Max tokens in one request (input + output) |
| Effort | low/medium/high/xhigh/max: how much effort/how many tokens |
| Embeddings | A numeric vector of text (for semantic search) |
| Eval | A test suite that measures LLM output quality |
| Few-shot | Giving examples in the prompt |
| Hallucination | Confident but wrong output |
| Hook | Deterministic code on a lifecycle event (Claude Code / Agent SDK) |
| IChatClient | The common chat interface of .NET Microsoft.Extensions.AI |
| ITPM / OTPM / RPM | Input tokens / output tokens / requests per minute limits |
| Managed Agents | Anthropic-hosted agent loop + sandbox |
| MCP | Model Context Protocol: the standard between AI apps ↔ tools/data |
| Memory tool | Remembering via client-side /memories files |
| Plan mode | Claude Code's read-only planning mode |
| Plugin | A bundle of commands + agents + skills + hooks + MCP |
| Prefill | Starting the assistant turn yourself; removed on newer models |
| Prompt caching | Making a repeated prefix cheaper and faster |
| Prompt injection | Malicious instructions hidden in data |
| RAG | Retrieve relevant docs → into the prompt → grounded answer |
| Refusal | stop_reason: refusal: the safety classifier declined |
| RLHF / RLAIF | Reinforcement learning from human / AI feedback |
| RSP | Responsible Scaling Policy |
| Skill | SKILL.md + files; procedural knowledge loaded on demand |
| SSE | Server-Sent Events: the streaming format |
| stop_reason | Why the response stopped (end_turn, max_tokens, tool_use...) |
| Structured outputs | A guaranteed format via JSON schema |
| Subagent | A specialized worker agent with its own context window |
| System prompt | A top-level instruction that sets role/rules |
| Token | A small unit of text; billing and limits are based on it |
| Tool Runner | An SDK helper that runs the tool-use loop |
| Tool use | Claude calling functions/APIs |
| Vault | Secrets in Managed Agents that are not visible in the sandbox |
| ZDR | Zero Data Retention |