Chapter 10: Last-Week Review

Summary: This chapter compresses the notes into decision rules and the distinctions exam items most often turn on. Use it to check yourself, not to learn things for the first time. If a line here surprises you, go back to the chapter it points to.

Ten decision rules that settle most scenarios

"Must never" or "must always" means enforcement. Use hooks, deny rules, tool lists, schemas or code checks, not prompt wording (Chapter 5).¹
Start simple. Use a single call, then a workflow, and an agent only when the steps cannot be known in advance (Chapter 3).²
A subagent knows only its prompt string. Every needed fact, path, prior finding and source goes in the handoff (Chapter 2).³
Restrict with tools, not with allowedTools. Allow lists pre-approve. They do not remove (Chapters 2 and 6).<sup>[3][4]</sup>
Answer all parallel tool calls in one user message. Every tool_use gets a tool_result (Chapter 1).⁵
Schemas guarantee shape, code checks truth. Validate, then retry with the specific error, then send to a person (Chapter 8).
Let schemas say "not found". Use nullable fields, enums with "other", and evidence fields (Chapter 8).⁶
Persist state outside the context. Use manifests, notes files and references. Resume from state, not from a stale transcript (Chapter 4).⁷
Review with fresh eyes. Use a separate pass or session with criteria, read-only tools and structured output (Chapter 6).⁸
Route review by risk. Use validation failures, missing fields, disagreement and stakes, not random sampling (Chapter 9).

Distinctions that decide close answers

Pair	The difference
`tools` vs `allowedTools`	`tools` defines what exists. `allowedTools` pre-approves<sup>[3][4]</sup>
Skill `allowed-tools` vs subagent `tools`	The first pre-approves during the skill. The second restricts⁹
Fork subagent vs `context: fork` skill	A fork inherits the conversation. A `context: fork` skill runs isolated, without history<sup>[3][9]</sup>
`--continue` vs `--resume` vs `--fork-session`	Latest session, a named session, or a branch that leaves the original unchanged¹⁰
CLAUDE.md vs `.claude/rules/` with `paths`	Always loaded vs loaded when matching files are touched¹¹
`@import` vs path rule	An import still loads at launch. A path rule defers loading¹¹
User vs project vs local MCP scope	All projects (private), shared `.mcp.json` (approved), this project (private, the default). Local wins¹²
MCP tool vs resource	Model-controlled action vs application-provided data¹³
Protocol error vs `isError` result	JSON-RPC failure vs an execution failure the model can fix¹⁴
Structured outputs vs strict tool use	Constrains the response vs constrains tool inputs. They work together⁶
`tool_choice` `any` vs `tool`	Some tool vs a named tool. Both skip preamble and thinking. Rejected on the newest models¹⁵
Messages vs Message Batches	Synchronous vs asynchronous at 50% cost, within 24 hours, matched by `custom_id`¹⁶
Compaction vs context editing	Summarize old turns vs clear used tool results or thinking server side¹⁷
Plan mode vs direct execution	Propose with edits blocked until approval vs act at once¹⁸
`--tools` vs `--allowedTools` in CI	Restrict built-in tools vs pre-approve them⁴

Numbers worth knowing

Fact	Value
Subagent spawn depth default	3 (`CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH`)³
Concurrent subagents default	20³
CLAUDE.md import depth	4 hops¹¹
Suggested CLAUDE.md length	Under about 200 lines¹
Few-shot examples	3 to 5¹⁹
Query at the end of long inputs	Up to 30% better in Anthropic's tests¹⁹
Batch limits	100,000 requests or 256 MB; 24-hour expiry; 29-day results; 50% price¹⁶
Server tool iteration default	10, then `pause_turn`²⁰
Claude Code tool response cap	25,000 tokens by default²¹
Multi-agent vs single agent (research eval)	90.2% better; about 15 times the tokens of chat²²

Model-specific facts that change answers

Forced tool_choice (any or tool) returns a 400 on Opus 5.5, Sonnet 5.5, Fable 5.1 and Mythos 5.1, and whenever manual extended thinking is enabled. Use auto with strict: true, or structured outputs.¹⁵
Prefilling the final assistant turn returns a 400 from Claude 4.6 models onward. Use structured outputs or direct instructions.¹⁹
Parallel calls on Fable 5.1 may drop off in long loops. Repeat the instruction after each round of tool results.¹⁹

A final self-check

Answer each in a sentence before the exam.

A coordinator spawns three researchers that keep covering the same sources. What is missing from the handoff?
A CI reviewer with --allowedTools "Read,Grep" still edits files. Why, and what is the fix?
A pipeline crashed at item 412 of 900. How does it resume without redoing work?
An extraction returns plausible due dates for invoices that have none. Which layer is wrong?
Three tool results come back in three user messages, and Claude stops making parallel calls. Why?
A migration rule in the root CLAUDE.md is ignored, and the file is 600 lines long. Where should the rule go?
A team commits .mcp.json with a token in it. What should it contain instead?
Human review uses a 5% random sample, and errors cluster in amended contracts. What should drive routing?
An Opus 5.5 integration sets tool_choice: {"type": "any"} and gets a 400. What is the replacement?
A long session forgets a rule that lived in a path-scoped rule file. What happened?

Answers, briefly: (1) objective, output format, source guidance and boundaries. (2) allowedTools only pre-approves, so use --tools or --disallowedTools. (3) A manifest with per-item status, and outputs written to files. (4) The schema forced a value, so make the field nullable. (5) Results must share one user message. (6) A paths:-scoped rule for the migrations directory. (7) ${VAR} expansion. (8) Conflict flags, null fields, failed quote checks, disagreement and stakes. (9) auto with strict tools, or structured outputs. (10) Compaction summarized it away, so make it unscoped or move it into the root CLAUDE.md.

Endnotes for this chapter

Deterministic controls are separated from advisory ones, and CLAUDE.md should stay under about 200 lines. Anthropic, "Best practices for Claude Code" (Claude Code Docs, accessed 2026-11-01), https://code.claude.com/docs/en/best-practices
Start with the simplest solution and add agentic complexity only when it is justified. Anthropic, "Building effective agents" (Anthropic Engineering, accessed 2026-11-01), https://www.anthropic.com/engineering/building-effective-agents
Subagents receive only their prompt string from the parent, are restricted by their tools list, can be forks, and are limited by spawn depth and concurrency settings. Anthropic, "Subagents in the SDK" (Claude Code Docs, accessed 2026-11-01), https://code.claude.com/docs/en/agent-sdk/subagents
--tools restricts built-in tools while --allowedTools lets listed tools run without prompts. Anthropic, "CLI reference" (Claude Code Docs, accessed 2026-11-01), https://code.claude.com/docs/en/cli-reference
All results for parallel calls belong in one user message. Anthropic, "Parallel tool use" (Claude Platform Docs, accessed 2026-11-01), https://platform.claude.com/docs/en/agents-and-tools/tool-use/parallel-tool-use
Structured outputs and strict tool use constrain different things and can be combined. Anthropic, "Structured outputs" (Claude Platform Docs, accessed 2026-11-01), https://platform.claude.com/docs/en/build-with-claude/structured-outputs
Capturing results as application state for a fresh session is often more robust than resuming. Anthropic, "Work with sessions" (Claude Code Docs, accessed 2026-11-01), https://code.claude.com/docs/en/agent-sdk/sessions
Independent review uses a fresh subagent or second session that did not write the code. Anthropic, "Best practices for Claude Code" (Claude Code Docs, accessed 2026-11-01), https://code.claude.com/docs/en/best-practices
A skill's allowed-tools pre-approves tools, and context: fork runs the skill in an isolated subagent. Anthropic, "Extend Claude with skills" (Claude Code Docs, accessed 2026-11-01), https://code.claude.com/docs/en/skills
Sessions can be continued, resumed by ID or name, or forked. Anthropic, "Manage sessions" (Claude Code Docs, accessed 2026-11-01), https://code.claude.com/docs/en/sessions
CLAUDE.md imports nest up to four hops and still load at launch, while path-scoped rules load when matching files are touched. Anthropic, "How Claude remembers your project" (Claude Code Docs, accessed 2026-11-01), https://code.claude.com/docs/en/memory
MCP servers can be configured in local, project or user scope, with local taking precedence. Anthropic, "Connect Claude Code to tools via MCP" (Claude Code Docs, accessed 2026-11-01), https://code.claude.com/docs/en/mcp
MCP tools are model-controlled and resources are application-driven. Anthropic, "Understanding MCP servers" (MCP Specification, accessed 2026-11-01), https://modelcontextprotocol.io/docs/2026-07-28/learn/server-concepts
MCP separates protocol errors from tool execution errors reported with isError. Anthropic, "Tools" (MCP Specification, accessed 2026-11-01), https://modelcontextprotocol.io/specification/2026-07-28/server/tools
Forced tool choice skips preamble and thinking and is rejected by the newest models and by manual extended thinking. Anthropic, "Define tools" (Claude Platform Docs, accessed 2026-11-01), https://platform.claude.com/docs/en/agents-and-tools/tool-use/define-tools
Message Batches cost half price, expire after 24 hours, keep results for 29 days and return results in any order. Anthropic, "Batch processing" (Claude Platform Docs, accessed 2026-11-01), https://platform.claude.com/docs/en/build-with-claude/batch-processing
Context editing clears tool results or thinking blocks server side, separate from summary-based compaction. Anthropic, "Context editing" (Claude Platform Docs, accessed 2026-11-01), https://platform.claude.com/docs/en/build-with-claude/context-editing
Plan mode blocks source edits until a plan is approved. Anthropic, "Choose a permission mode" (Claude Code Docs, accessed 2026-11-01), https://code.claude.com/docs/en/permission-modes
The prompting guide recommends 3 to 5 examples and queries at the end of long inputs, says prefill returns a 400 from Claude 4.6 onward, and covers parallel calling on Fable 5.1. Anthropic, "Prompting best practices" (Claude Platform Docs, accessed 2026-11-01), https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practices
Server tools pause with pause_turn after their default iteration limit. Anthropic, "How tool use works" (Claude Platform Docs, accessed 2026-11-01), https://platform.claude.com/docs/en/agents-and-tools/tool-use/how-tool-use-works
Claude Code limits tool responses to 25,000 tokens by default. Anthropic, "Writing effective tools for agents — with agents" (Anthropic Engineering, accessed 2026-11-01), https://www.anthropic.com/engineering/writing-tools-for-agents
The multi-agent research system outperformed a single agent by 90.2% and used about 15 times the tokens of chat. Anthropic, "How we built our multi-agent research system" (Anthropic Engineering, accessed 2026-11-01), https://www.anthropic.com/research/multi-agent-research-system

Ten decision rules that settle most scenarios

"Must never" or "must always" means enforcement. Use hooks, deny rules, tool lists, schemas or code checks, not prompt wording (Chapter 5).¹

Start simple. Use a single call, then a workflow, and an agent only when the steps cannot be known in advance (Chapter 3).²

A subagent knows only its prompt string. Every needed fact, path, prior finding and source goes in the handoff (Chapter 2).³

Restrict with tools, not with allowedTools. Allow lists pre-approve. They do not remove (Chapters 2 and 6).<sup>[3][4]</sup>

Answer all parallel tool calls in one user message. Every tool_use gets a tool_result (Chapter 1).⁵

Schemas guarantee shape, code checks truth. Validate, then retry with the specific error, then send to a person (Chapter 8).

Let schemas say "not found". Use nullable fields, enums with "other", and evidence fields (Chapter 8).⁶

Persist state outside the context. Use manifests, notes files and references. Resume from state, not from a stale transcript (Chapter 4).⁷

Review with fresh eyes. Use a separate pass or session with criteria, read-only tools and structured output (Chapter 6).⁸

Route review by risk. Use validation failures, missing fields, disagreement and stakes, not random sampling (Chapter 9).

Distinctions that decide close answers

Pair	The difference
`tools` vs `allowedTools`	`tools` defines what exists. `allowedTools` pre-approves<sup>[3][4]</sup>
Skill `allowed-tools` vs subagent `tools`	The first pre-approves during the skill. The second restricts⁹
Fork subagent vs `context: fork` skill	A fork inherits the conversation. A `context: fork` skill runs isolated, without history<sup>[3][9]</sup>
`--continue` vs `--resume` vs `--fork-session`	Latest session, a named session, or a branch that leaves the original unchanged¹⁰
CLAUDE.md vs `.claude/rules/` with `paths`	Always loaded vs loaded when matching files are touched¹¹
`@import` vs path rule	An import still loads at launch. A path rule defers loading¹¹
User vs project vs local MCP scope	All projects (private), shared `.mcp.json` (approved), this project (private, the default). Local wins¹²
MCP tool vs resource	Model-controlled action vs application-provided data¹³
Protocol error vs `isError` result	JSON-RPC failure vs an execution failure the model can fix¹⁴
Structured outputs vs strict tool use	Constrains the response vs constrains tool inputs. They work together⁶
`tool_choice` `any` vs `tool`	Some tool vs a named tool. Both skip preamble and thinking. Rejected on the newest models¹⁵
Messages vs Message Batches	Synchronous vs asynchronous at 50% cost, within 24 hours, matched by `custom_id`¹⁶
Compaction vs context editing	Summarize old turns vs clear used tool results or thinking server side¹⁷
Plan mode vs direct execution	Propose with edits blocked until approval vs act at once¹⁸
`--tools` vs `--allowedTools` in CI	Restrict built-in tools vs pre-approve them⁴

Numbers worth knowing

Fact	Value
Subagent spawn depth default	3 (`CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH`)³
Concurrent subagents default	20³
CLAUDE.md import depth	4 hops¹¹
Suggested CLAUDE.md length	Under about 200 lines¹
Few-shot examples	3 to 5¹⁹
Query at the end of long inputs	Up to 30% better in Anthropic's tests¹⁹
Batch limits	100,000 requests or 256 MB; 24-hour expiry; 29-day results; 50% price¹⁶
Server tool iteration default	10, then `pause_turn`²⁰
Claude Code tool response cap	25,000 tokens by default²¹
Multi-agent vs single agent (research eval)	90.2% better; about 15 times the tokens of chat²²

Model-specific facts that change answers

Forced tool_choice (any or tool) returns a 400 on Opus 5.5, Sonnet 5.5, Fable 5.1 and Mythos 5.1, and whenever manual extended thinking is enabled. Use auto with strict: true, or structured outputs.¹⁵

Prefilling the final assistant turn returns a 400 from Claude 4.6 models onward. Use structured outputs or direct instructions.¹⁹

Parallel calls on Fable 5.1 may drop off in long loops. Repeat the instruction after each round of tool results.¹⁹

A final self-check

Answer each in a sentence before the exam.

A coordinator spawns three researchers that keep covering the same sources. What is missing from the handoff?

A CI reviewer with --allowedTools "Read,Grep" still edits files. Why, and what is the fix?

A pipeline crashed at item 412 of 900. How does it resume without redoing work?

An extraction returns plausible due dates for invoices that have none. Which layer is wrong?

Three tool results come back in three user messages, and Claude stops making parallel calls. Why?

A migration rule in the root CLAUDE.md is ignored, and the file is 600 lines long. Where should the rule go?

A team commits .mcp.json with a token in it. What should it contain instead?

Human review uses a 5% random sample, and errors cluster in amended contracts. What should drive routing?

An Opus 5.5 integration sets tool_choice: {"type": "any"} and gets a 400. What is the replacement?

A long session forgets a rule that lived in a path-scoped rule file. What happened?

Endnotes for this chapter

Deterministic controls are separated from advisory ones, and CLAUDE.md should stay under about 200 lines. Anthropic, "Best practices for Claude Code" (Claude Code Docs, accessed 2026-11-01), https://code.claude.com/docs/en/best-practices

Start with the simplest solution and add agentic complexity only when it is justified. Anthropic, "Building effective agents" (Anthropic Engineering, accessed 2026-11-01), https://www.anthropic.com/engineering/building-effective-agents

Subagents receive only their prompt string from the parent, are restricted by their tools list, can be forks, and are limited by spawn depth and concurrency settings. Anthropic, "Subagents in the SDK" (Claude Code Docs, accessed 2026-11-01), https://code.claude.com/docs/en/agent-sdk/subagents

--tools restricts built-in tools while --allowedTools lets listed tools run without prompts. Anthropic, "CLI reference" (Claude Code Docs, accessed 2026-11-01), https://code.claude.com/docs/en/cli-reference

All results for parallel calls belong in one user message. Anthropic, "Parallel tool use" (Claude Platform Docs, accessed 2026-11-01), https://platform.claude.com/docs/en/agents-and-tools/tool-use/parallel-tool-use

Structured outputs and strict tool use constrain different things and can be combined. Anthropic, "Structured outputs" (Claude Platform Docs, accessed 2026-11-01), https://platform.claude.com/docs/en/build-with-claude/structured-outputs

Capturing results as application state for a fresh session is often more robust than resuming. Anthropic, "Work with sessions" (Claude Code Docs, accessed 2026-11-01), https://code.claude.com/docs/en/agent-sdk/sessions

Independent review uses a fresh subagent or second session that did not write the code. Anthropic, "Best practices for Claude Code" (Claude Code Docs, accessed 2026-11-01), https://code.claude.com/docs/en/best-practices

A skill's allowed-tools pre-approves tools, and context: fork runs the skill in an isolated subagent. Anthropic, "Extend Claude with skills" (Claude Code Docs, accessed 2026-11-01), https://code.claude.com/docs/en/skills

Sessions can be continued, resumed by ID or name, or forked. Anthropic, "Manage sessions" (Claude Code Docs, accessed 2026-11-01), https://code.claude.com/docs/en/sessions

CLAUDE.md imports nest up to four hops and still load at launch, while path-scoped rules load when matching files are touched. Anthropic, "How Claude remembers your project" (Claude Code Docs, accessed 2026-11-01), https://code.claude.com/docs/en/memory

MCP servers can be configured in local, project or user scope, with local taking precedence. Anthropic, "Connect Claude Code to tools via MCP" (Claude Code Docs, accessed 2026-11-01), https://code.claude.com/docs/en/mcp

MCP tools are model-controlled and resources are application-driven. Anthropic, "Understanding MCP servers" (MCP Specification, accessed 2026-11-01), https://modelcontextprotocol.io/docs/2026-07-28/learn/server-concepts

MCP separates protocol errors from tool execution errors reported with isError. Anthropic, "Tools" (MCP Specification, accessed 2026-11-01), https://modelcontextprotocol.io/specification/2026-07-28/server/tools

Forced tool choice skips preamble and thinking and is rejected by the newest models and by manual extended thinking. Anthropic, "Define tools" (Claude Platform Docs, accessed 2026-11-01), https://platform.claude.com/docs/en/agents-and-tools/tool-use/define-tools

Message Batches cost half price, expire after 24 hours, keep results for 29 days and return results in any order. Anthropic, "Batch processing" (Claude Platform Docs, accessed 2026-11-01), https://platform.claude.com/docs/en/build-with-claude/batch-processing

Context editing clears tool results or thinking blocks server side, separate from summary-based compaction. Anthropic, "Context editing" (Claude Platform Docs, accessed 2026-11-01), https://platform.claude.com/docs/en/build-with-claude/context-editing

Plan mode blocks source edits until a plan is approved. Anthropic, "Choose a permission mode" (Claude Code Docs, accessed 2026-11-01), https://code.claude.com/docs/en/permission-modes

The prompting guide recommends 3 to 5 examples and queries at the end of long inputs, says prefill returns a 400 from Claude 4.6 onward, and covers parallel calling on Fable 5.1. Anthropic, "Prompting best practices" (Claude Platform Docs, accessed 2026-11-01), https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practices

Server tools pause with pause_turn after their default iteration limit. Anthropic, "How tool use works" (Claude Platform Docs, accessed 2026-11-01), https://platform.claude.com/docs/en/agents-and-tools/tool-use/how-tool-use-works

Claude Code limits tool responses to 25,000 tokens by default. Anthropic, "Writing effective tools for agents — with agents" (Anthropic Engineering, accessed 2026-11-01), https://www.anthropic.com/engineering/writing-tools-for-agents

The multi-agent research system outperformed a single agent by 90.2% and used about 15 times the tokens of chat. Anthropic, "How we built our multi-agent research system" (Anthropic Engineering, accessed 2026-11-01), https://www.anthropic.com/research/multi-agent-research-system

Chapter 10: Last-Week Review

Ten decision rules that settle most scenarios

Distinctions that decide close answers

Numbers worth knowing

Model-specific facts that change answers

A final self-check

Endnotes for this chapter

For detailed info, read the official docs

Chapter 10: Last-Week Review

Ten decision rules that settle most scenarios

Distinctions that decide close answers

Numbers worth knowing

Model-specific facts that change answers

A final self-check

Endnotes for this chapter

For detailed info, read the official docs