Glossary

Terms you should be able to define cold.

Agent / Workflow

• Workflow — pre-defined steps; control flow at design time.
• Agent — LLM dynamically decides next step at runtime.

Agentic patterns

• Prompt chaining — sequential LLM calls; each output feeds next.
• Routing — classifier dispatches to specialized handlers.
• Parallelization — independent calls run concurrently; results merged.
• Sectioning — split input into independent chunks.
• Voting — same prompt N times; majority/consensus.
• Orchestrator-workers — central LLM plans + delegates dynamically.
• Evaluator-optimizer — generator/critic loop until quality bar met.

Building blocks

• Reflection — model critiques its own output.
• Self-correction loop — validation failure feeds back to model.
• Provenance — source IDs threaded through results.
• Confirmation gate — human approves destructive action.

API concepts

• Messages API — POST /v1/messages. Stateless.
• Content block — text, tool_use, tool_result, thinking, image, document.
• stop_reason — end_turn, tool_use, max_tokens, stop_sequence, pause_turn.
• tool_choice — auto | any | {type: "tool", name} | none.
• Prefill — start assistant turn with text to constrain output.
• Extended thinking — model emits thinking blocks before answer.
• Signature — required metadata on thinking blocks for tool-use continuity.

Prompt-engineering vocabulary

• Few-shot — examples in the prompt.
• CoT (Chain of Thought) — "think step by step".
• Dilution — long prompt causes critical rule to be missed.
• Principles vs conditionals — universal rules vs if-then exceptions; principles preferred.

Caching

• cache_control: {type: "ephemeral"} — mark cacheable prefix.
• Cache hit — cache_read_input_tokens > 0 in usage.
• Breakpoint — boundary marking a cacheable unit.

MCP

• Host — application running clients (Claude Desktop, Claude Code).
• Client — connection to one MCP server.
• Server — exposes tools/resources/prompts over JSON-RPC.
• Tool — model-controlled action.
• Resource — application-controlled read-only data; identified by URI.
• Prompt — user-controlled parameterized template.
• Roots — directory scope the server can access.
• Sampling — server asks client to run an LLM call.
• Transports — stdio, SSE (legacy), Streamable HTTP.

Claude Code

• CLAUDE.md — durable memory file Claude reads automatically.
• Slash command — user-invoked prompt template.
• Skill — reusable capability bundle (SKILL.md + files).
• Subagent — isolated agent the main session delegates to.
• Hook — shell command fired on lifecycle events.
• Plan mode — read-only; agent proposes plan before executing.
• Permission — Tool(arg-glob) allow/deny entry.
• Worktree isolation — subagent works in a separate git worktree.

Cowork

• Standing context — instructions/materials Claude carries into every task (global instructions + projects).
• Background task execution — a Cowork session keeps running with no device online; output waits for you.
• Cross-device session sync — start on one device, monitor/finish on another.

Reliability

• Idempotent — repeating the call yields the same result.
• Transient / permanent / uncertain error types.
• Defensive read — check status before retrying non-idempotent op.
• Graceful degradation — useful fallback when tool fails.
• Escalation — hand off to human / higher-confidence path.

Model tiers

• Opus — most capable, most expensive.
• Sonnet — balanced default.
• Haiku — fastest, cheapest.

Frameworks / mnemonics

• CALM — Cache, Align, Limit, Monitor (context management).
• HHH — Helpful, Harmless, Honest.
• CAI / Constitutional AI — training Claude against a written constitution.

See also

• cheatsheet
• cca-f-exam-overview