CCA-F start here · Exam vs 2026

Exam vs 2026: the gotchas

The exam is fixed to its guide, but the Claude API and Claude Code keep changing. In a few places, the answer the exam wants is no longer how you'd build it today. If you know the current product well, those questions feel wrong. This page lists them, so you can recognise them and answer the way the exam expects.

All domainsChecked against the exam guide and the live API reference, 31 Oct 2026About 15 minutes

The one rule: answer from the exam guide. Use the newer way only when an option explicitly offers it and nothing else fits.

At a glance

#TopicExam answer (follow this)What's true today
1Guaranteed JSONA tool with a JSON schemaoutput_config.format (structured outputs)
2Forcing a tool calltool_choice any or forced are validBoth return an error on Opus 5.5, Sonnet 5.5 and Fable 5.1
3Spawning subagentsThe Task tool; allowedTools must include "Task"Claude Code calls it the Agent tool
4Custom slash commands.claude/commands/ (shared) vs ~/.claude/commands/ (personal)Commands are folded into skills
5MCP error for a bad argumentA JSON-RPC protocol errorA late-2025 spec revision suggests isError
6stop_reasonTwo values matter: tool_use and end_turnSeven values exist
7Too many toolsGive each agent only the 4 or 5 tools it needsTool search lets one agent have hundreds
8Prefill and temperaturePrefill is a wrong option; temperature is barely testedBoth return an error on current models

The eight, one by one

1 · Guaranteed JSON means “a tool with your schema”

Exam: guide 4.3 says “tool_use with JSON schemas = most reliable schema-compliant output”. You define a tool whose input schema is the shape you want. Claude “calls” it with the extracted values, and those inputs are your JSON. No code runs; the tool is a form Claude fills in.

Today: a dedicated feature, structured outputs (output_config.format), does the same thing without a pretend tool.

Why it's called a tool: the schema guarantee was first built for tool inputs, so that real functions wouldn't break. Developers borrowed it to get clean JSON, and the name stuck.

How to read it: “define a tool with your target schema” means the guaranteed-JSON option.

Example. Extracted invoice data must match a schema exactly, or the downstream system rejects it. Most reliable?
  1. Define a tool whose input schema is the invoice schema, and have Claude call it with the extracted values
  2. Prefill the reply with {
  3. Ask for “only valid JSON” and retry when parsing fails
  4. Put the schema in the prompt and parse the text
A. The API enforces the shape. The others only ask Claude to comply.

2 · tool_choice: any and forced are valid answers

Exam: auto means Claude may reply in text. any means it must call some tool. Forced ({"type":"tool","name":"..."}) means it must call that exact tool.

Today: any and forced return an error on Opus 5.5, Sonnet 5.5 and Fable 5.1. You use auto plus an instruction, or structured outputs.

How to read it: “must always produce the extraction” means forced. “Must use one of these tools but may pick” means any. “May answer in plain text” means auto.

Example. An extraction pipeline sometimes gets a text summary instead of a record_invoice call. Fix?
  1. Add “always use the tool” to the system prompt
  2. Set tool_choice to force record_invoice
  3. Set tool_choice: "auto"
  4. Retry when the reply has no tool call
B. Forcing guarantees the call. A is only a request, C is the cause of the problem, and D catches the problem after it happens.

3 · Subagents are spawned by the “Task” tool

Exam: guide 1.2 says the Task tool spawns subagents and allowedTools must include "Task". For parallel subagents, the coordinator emits several Task calls in one response.

Today: Claude Code calls this tool Agent.

How to read it: “Task” means the subagent launcher. If a coordinator never delegates, check whether "Task" is missing from allowedTools.

Example. A coordinator built with the Agent SDK describes delegating to subagents but never actually launches any. Most likely cause?
  1. allowedTools doesn't include "Task"
  2. The subagents' system prompts are too short
  3. max_tokens is too low
  4. The subagents need their own API keys
A. Without the Task tool, the coordinator can only talk about delegating.

4 · Shared slash commands live in .claude/commands/

Exam: project commands in .claude/commands/ are version-controlled and shared with the team. Personal ones go in ~/.claude/commands/. Skills (.claude/skills/ with SKILL.md) are on-demand workflows, and context: fork runs one in an isolated sub-agent.

Today: commands have been folded into skills, and a skill is also a slash command.

How to read it: “team-wide command, available to everyone who clones the repo” means .claude/commands/. “A workflow with its own allowed tools or an isolated context” means a skill.

Example. Every developer who clones the repo should get a /review command. Where?
  1. .claude/commands/ in the repo
  2. ~/.claude/commands/ on each machine
  3. Root CLAUDE.md
  4. A commands array in .claude/config.json
A. It's in the project and in git, so everyone gets it. B is personal, C is for context, and D doesn't exist.

5 · MCP: protocol error vs isError

Exam: if the call itself is broken (unknown tool, missing or invalid argument), the tool never ran, so it's a JSON-RPC protocol error. If the tool ran and its work failed (an API returned 404 or 503), it's a result with isError: true, which the model sees and can act on.

Today: a late-2025 spec revision recommends returning input-validation errors as isError too, so the model can correct its own call.

How to read it: “did the tool run?” If not, it's a protocol error. If it ran and failed, it's isError. The downstream HTTP code doesn't change that.

Example. Three failures: (1) user_email is missing, (2) the calendar API returns 404 because the user doesn't exist, (3) the calendar API returns 503. How should each be reported?
  1. All three as isError
  2. 1 and 2 as protocol errors, 3 as isError
  3. 1 as a protocol error, 2 and 3 as isError
  4. All three as protocol errors
C. Only 1 stopped the tool from running. B is the trap: a 404 sounds like a protocol error, but it came from the calendar API.

6 · stop_reason: only two values matter

Exam: keep looping while stop_reason == "tool_use", and stop at "end_turn".

Today: there are seven values, also including max_tokens, stop_sequence, pause_turn, refusal and model_context_window_exceeded.

How to read it: options that stop the loop when text appears, or after a fixed number of turns, are wrong. stop_reason controls the loop.

Example. An agent sometimes stops after its first tool call. The loop exits whenever the reply contains a text block. Fix?
  1. Raise the iteration limit
  2. Continue while stop_reason is tool_use, and stop on end_turn
  3. Tell Claude not to write text before tool calls
  4. Stop when the reply contains the word “done”
B. Text can come before a tool call. Only stop_reason says whether Claude has finished.

7 · Too many tools: scope them per agent

Exam: guide 2.3 says 18 tools degrade selection compared with 4 or 5. Give each subagent only the tools for its role.

Today: tool search (defer_loading) lets an agent have hundreds of tools and load only the ones it searches for.

How to read it: usually the answer is “give each subagent a few tools for its role”. Pick dynamic loading only when an option offers it and the problem is a large connector library.

Example. A single agent with 18 tools often chooses the wrong one. Best fix?
  1. Longer descriptions on all 18 tools
  2. Split the work across subagents, each with the 4 or 5 tools for its role
  3. Few-shot examples for every tool
  4. Merge the 18 tools into one generic tool
B. Fewer, role-specific tools per agent. A and C add text without reducing the choice, and D hides it.

8 · Prefill and temperature

Exam: prefilling { appears as a wrong option for reliable JSON, because it starts the reply as JSON but doesn't enforce your schema. Temperature is barely tested.

Today: prefill returns an error on current models, and temperature, top_p and top_k return an error on API version 2026-10-21.

How to read it: the exam and today agree here, so never choose prefill as the reliable answer. If an old question uses temperature for consistency, prefer options that change structure (a schema, examples, explicit criteria) over a sampling setting.

Where instinct points the wrong way

These don't depend on product versions. They're habits that feel right and lose marks.

#InstinctExam answerPractise
9“Claude can check its own work if it thinks harder”An independent review instance with a fresh context and no earlier reasoning catches moreSelf-review
10“Retry until it's valid”Retry with the specific error. It's useless when the information isn't in the source; use a nullable field or human reviewExtraction honesty
11“A strict schema means correct data”A schema fixes syntax, never meaning (totals that don't add up)Extraction honesty
12“Batch everything to save 50%”Batch only work that can wait up to 24 hours with no time guarantee. Nothing that blocks a person or a merge, and no multi-turn tool callingBatch or now
13“A hook is code, so it's safe”Only a hook that blocks is enforcement. A hook that adds a flag or logs is not. For tamper-proof rules, put the check inside the toolPrompt or hook: tool A before tool B
14“Code beats prompts, so always pick code”Only when the question asks for a guarantee. For “improve tool selection” with vague descriptions, better descriptions winTool picker
15“One CLAUDE.md per folder”Conventions that cross folders (like **/*.test.tsx) go in .claude/rules/ with glob patternsWhere an instruction lives

Rule to remember

The exam is fixed to its guide. When today's product disagrees, answer from the guide: a tool with a schema for JSON, tool_choice forced or any, the Task tool, .claude/commands/, protocol error vs isError, and stop_reason controlling the loop. Use the newer way only when an option explicitly offers it and nothing else fits.

Related: every API field, with what to study · all traps · CCA-F start here