Every concept from the three guides, deduplicated and sorted alphabetically. Filter by guide, shuffle the deck, or flip everything at once. Five cards appear in multiple guides — they're tagged with all the guides they cover.
Showing 45 cards · 0 flipped
01 · Foundations
Action
tap to reveal →
The step the model chooses to take — a bash command, a JSON tool call, or a Python snippet. Executed by the environment.
← tap to flip back
01 · Foundations02 · Applied
Agent loop
tap to reveal →
LLM decides → tool executes → result feeds back, repeated until the goal is met. In Claude Code framing: gather context → take action → verify → repeat. The core of every agent.
← tap to flip back
02 · Applied
Agent SDK
tap to reveal →
Library to embed Claude Code's loop in your own apps with full control over tools, permissions, and orchestration logic.
← tap to flip back
02 · Applied
Agent team
tap to reveal →
Multiple full agents on parallel workstreams — e.g. one on backend, one on frontend tests — coordinated by a lead agent that merges results.
← tap to flip back
02 · Applied
Agentic harness
tap to reveal →
Everything around the model — tools, context management, execution environment — that turns an LLM into a coding agent.
← tap to flip back
02 · Applied
Auto memory
tap to reveal →
Learnings Claude saves itself (to MEMORY.md) and reloads across sessions — build commands, debugging insights — no writing required.
← tap to flip back
02 · Applied
Checkpoint
tap to reveal →
A pre-edit file snapshot taken automatically. Press Esc Esc to rewind. Local, separate from git; applies to file changes only.
← tap to flip back
02 · Applied
CLAUDE.md
tap to reveal →
Project-root markdown file of standards, commands, and context — read at the start of every session. Use /init to generate a starter.
← tap to flip back
01 · Foundations
CodeAgent
tap to reveal →
smolagents' default agent type: writes actions as executable Python run in a sandbox. Composable — loops, conditionals, and nesting in one step.
← tap to flip back
01 · Foundations02 · Applied
Compaction
tap to reveal →
Auto-summarizing or dropping older turns when the context window fills, to keep the goal and key results while shedding stale detail. Steerable via /compact in Claude Code.
← tap to flip back
03 · Advanced
Continuous batching
tap to reveal →
Token-level dynamic batching of concurrent users' requests to maximize GPU throughput. Standard practice in production LLM serving.
← tap to flip back
03 · Advanced
Dreaming / nightly tide
tap to reveal →
Offline batch consolidation of memory — scheduled nightly (T+1) instead of real-time writes. Avoids the hard problem of deciding what to write mid-conversation.
← tap to flip back
01 · Foundations
Function calling
tap to reveal →
Model emits a structured tool call directly; deterministic and predictable. Reasoning happens inside the model, invisible in the output.
← tap to flip back
03 · Advanced
Goal drift
tap to reveal →
Long-horizon failure where the agent gradually loses or mutates its original objective — often subtle, compounding over many steps.
← tap to flip back
03 · Advanced
Harness
tap to reveal →
The engineering shell around the model — tools, context management, sandbox, loop control — everything that is not the model weights. Makes an agent reliable.
← tap to flip back
02 · Applied
Hook
tap to reveal →
A shell command that runs automatically before or after Claude's actions — format-on-edit, lint-before-commit, type-check after any file change.
← tap to flip back
03 · Advanced
Human-in-the-loop (HITL)
tap to reveal →
A gate where the loop pauses for human approval or input before continuing — used for risky or irreversible actions.
← tap to flip back
03 · Advanced
KV cache (key-value cache)
tap to reveal →
Cached attention data for an unchanged prompt prefix. A cache hit avoids expensive recomputation. Prefix stability = cache hits = lower cost and latency.
← tap to flip back
01 · Foundations
LangChain / CrewAI / AutoGen
tap to reveal →
Heavyweight orchestration frameworks. Useful convenience layers once you understand the loop — but starting here hides the fundamentals.
← tap to flip back
01 · Foundations
Linear history
tap to reveal →
Every step just appends to the message list — nothing hidden, so runs are easy to replay and debug step by step.
← tap to flip back
01 · Foundations02 · Applied
MCP
tap to reveal →
Model Context Protocol — open standard to connect external data/tools (Jira, Drive, Slack, databases, files) as native tools without bespoke integration glue per service.
← tap to flip back
01 · Foundations
Memory vs. context
tap to reveal →
Context = live window (RAM) — finite, gone between sessions. Memory = persistence beyond the window (disk) — e.g. a Session Note loaded back in when relevant.
← tap to flip back
03 · Advanced
MicroVM
tap to reveal →
Lightweight virtual machine (Firecracker-style) — full VM isolation with near-container startup speed. Used for agent sandboxes to safely run arbitrary code.
← tap to flip back
01 · Foundations
Model / env / agent split
tap to reveal →
The clean 3-way separation: which LLM (model), where actions run (environment), and the loop logic (agent) — each replaceable independently.
← tap to flip back
01 · Foundations
MultiStepAgent
tap to reveal →
smolagents' base class implementing the ReAct loop. CodeAgent and ToolCallingAgent both extend it.
← tap to flip back
02 · Applied
Non-interactive mode
tap to reveal →
claude -p "…" — pipes Claude into CI, hooks, and scripts. Use --output-format json for programmatic parsing of results.
← tap to flip back
01 · Foundations
Observation
tap to reveal →
The result of executing an action, appended back into context so the model can reason from it on the next step.
← tap to flip back
02 · Applied
Plan mode
tap to reveal →
Read-only permission mode — Claude produces a plan for you to approve before making any change. Press Shift+Tab twice to activate.
← tap to flip back
03 · Advanced
Prefill / Decode
tap to reveal →
Prefill = parallel prompt processing (compute-bound, fast). Decode = token-by-token answer generation (memory-bandwidth-bound, slower). Long answers are the bottleneck.
← tap to flip back
03 · Advanced
Prefill intervention
tap to reveal →
Steering output by pre-writing the opening tokens of the model's response — strongly nudges the direction or format of what comes next.
← tap to flip back
03 · Advanced
Process reward model (PRM)
tap to reveal →
A model that gives step-by-step reward signals during RL training — much denser feedback for long agentic tasks than a final-outcome reward alone.
← tap to flip back
03 · Advanced
Progressive disclosure
tap to reveal →
Only keep a skill's name and description in context at startup (~60 tokens); load the full body on-demand when triggered. Saves tens of thousands of tokens.
← tap to flip back
01 · Foundations
ReAct
tap to reveal →
Reasoning + Acting: the Think → Act → Observe cycle, with reasoning made explicit in text before each action.
← tap to flip back
02 · Applied
Routines
tap to reveal →
Scheduled recurring tasks on Anthropic-managed infra — morning PR reviews, overnight CI-failure analysis — run even when your machine is off.
← tap to flip back
01 · Foundations
Sandbox
tap to reveal →
Isolated environment for running an agent's commands/code safely, away from the host — so mistakes can't do real damage.
← tap to flip back
01 · Foundations
Session Note
tap to reveal →
A persisted record of a run's learnings, saved to disk and reloaded in future sessions — one of the simplest forms of long-term memory.
← tap to flip back
01 · Foundations02 · Applied
Skill
tap to reveal →
A packaged, on-demand workflow (instructions + optional scripts) the agent loads only when relevant — e.g. /review-pr, /deploy-staging. Keeps context clean until needed.
← tap to flip back
03 · Advanced
Snapshot / Resume
tap to reveal →
Snapshot = save full machine state. Resume = restore instantly. Together they hide cold-start latency for sandboxes without sacrificing isolation.
← tap to flip back
03 · Advanced
Stop reason
tap to reveal →
Why generation halted: turn finished, model wants a tool call, or hit max tokens. The harness branches on this signal to decide what to do next.
← tap to flip back
02 · Applied
Subagent
tap to reveal →
A disposable helper with its own context window. Does one scoped job and returns a summary — keeps deep digging from clogging the main session.
← tap to flip back
01 · Foundations
subprocess.run
tap to reveal →
How mini-swe-agent executes each bash action independently — stateless, no persistent shell session, simple to reason about.
← tap to flip back
01 · Foundations03 · Advanced
SWE-bench
tap to reveal →
Benchmark of real GitHub issues — the agent must produce a patch that makes existing tests pass. Score reflects both model and harness quality. mini-swe-agent scores >74% on the Verified split.
← tap to flip back
01 · Foundations
Tool
tap to reveal →
A named, described function the model can invoke (web search, file read, API call) with typed inputs and outputs.
← tap to flip back
01 · Foundations
ToolCallingAgent
tap to reveal →
smolagents variant that writes actions as JSON tool calls — the classic function-calling style. Predictable and easy to govern.
← tap to flip back
02 · Applied
Worktree
tap to reveal →
An isolated git checkout so parallel Claude sessions don't collide on edits. The Desktop app manages these visually.