03 — Quiz: context engineering

Pick an answer. The quiz checks it and shows the explanation.

Set A

#

1. Which is loaded IN FULL at session start?

a) every skill body

b) root CLAUDE.md

c) every MCP tool schema

d) all path-scoped rules

b — skills and MCP tools load only a label at start. Path-scoped rules wait for a matching file.

#

2. Where does a 40-step weekly release procedure belong?

a) root CLAUDE.md

b) MEMORY.md

c) a skill

d) a path-scoped rule

c — a skill. Only its name and description load at start. The 40 steps load when you invoke it.

#

3. You said "staging is eu-west-3" in chat 3 hours ago. After compaction, Claude forgot it. Best fix?

a) repeat it

b) a path-scoped rule

c) root CLAUDE.md

d) raise MAX_MCP_OUTPUT_TOKENS

c — root CLAUDE.md is re-injected from disk after compaction. Repeating it in chat can be lost again.

#

4. What happens to a path-scoped rule for migrations/** after compaction?

a) reloaded immediately

b) summarized away; reloads on the next matching read

c) gone for the rest of the session

d) saved to memory

b — it comes back when Claude reads a file under migrations/ again.

#

5. What does a subagent see at start?

a) your full conversation

b) nothing, not even CLAUDE.md

c) the last tool result

d) the task message + standard material like CLAUDE.md

d — a fresh window with the task and standard material. Not your conversation. (The built-in Explore and Plan subagents even skip CLAUDE.md.)

#

6. A subagent reads 40 files and returns a 6,000-token report. What does it cost the main context?

a) ~6,000 tokens

b) zero

c) all 40 files

d) nothing until compaction

a — the 40 files stay in the subagent's window. Only the report lands in yours.

#

7. What does a PreToolUse hook cost in context?

a) the script is loaded at start

b) zero, unless it outputs content that gets added

c) it loads with skills

d) it is summarized at compaction

b — a hook is code outside the model. Only context it adds to the conversation costs tokens.

#

8. Claude must NEVER run terraform destroy in prod. Where do you put that?

a) CLAUDE.md

b) a skill

c) MEMORY.md

d) a PreToolUse hook or a deny rule

d — the other three are advisory context. Only a hook or a deny rule is enforced.

#

9. When should you run /compact manually?

a) mid-debugging at 50%

b) after every tool call

c) between tasks

d) never

c — between tasks. Mid-task, you may lose details you still need.

#

10. You have 100 skills of ~3,000 tokens each. What is the startup cost?

a) ~300,000 tokens

b) names + descriptions only

c) nothing

d) the last used skill

b — label upfront, body on demand.

Set B

#

1. You read 25 config files an hour ago. Answers now feel less focused. Why?

a) CLAUDE.md grows

b) the config files are still in context

c) skills reload

d) MCP re-runs tools/list

b — tool results stay in context until compaction. /compact or a subagent would have helped.

#

2. Which costs context every session, even if unused?

a) a 2,000-token skill body

b) a path-scoped rule

c) a 500-line root CLAUDE.md

d) subagent work

c — root CLAUDE.md loads in full at every start.

#

3. Your CLAUDE.md is 600 lines and its rules get ignored. Fix?

a) split it with @imports

b) write rules in CAPITALS

c) move it to MEMORY.md

d) keep it under ~200 lines; procedures → skills; file rules → path-scoped rules

d — @imports still load in full, so (a) changes nothing.

#

4. CLAUDE.md contains @docs/runbook.md (2,000 lines). What happens?

a) it loads in full every session

b) it loads when relevant

c) the first 200 lines load

d) it is never loaded

a — @imports are not lazy.

#

5. MEMORY.md is 400 lines. What loads at start?

a) all of it

b) nothing

c) the first 200 lines or 25KB

d) topic files only

c — keep MEMORY.md a short index. Topic files load on demand.

#

6. You spend 30 minutes explaining a bug, then ask a subagent to "look into the logs". The result is generic. Why?

a) it uses a weaker model

b) it never saw your conversation, only the vague task

c) it can't read files

d) it was compacted

b — put the facts it needs into the task message.

#

7. What is most likely missing after compaction?

a) a chat-only instruction

b) a root CLAUDE.md rule

c) an unscoped rule

d) the MEMORY.md index

a — the other three are reloaded from disk.

#

8. A hook that runs after every edit sends a 3,000-line lint report back to Claude. Context cost?

a) zero

b) the script, once

c) the first edit only

d) the report is added each time, so it grows fast

d — hooks are free only when they add nothing. Send a short summary instead.

#

9. Migration rules that apply only to migrations/. Where?

a) root CLAUDE.md

b) a path-scoped rule

c) a deny rule

d) a subagent

b — it loads only when Claude reads or edits a matching file.

#

10. You need to scan 50 log files and keep the main context clean. Best approach?

a) the main session reads them all

b) /compact, then read them

c) a subagent with "return only the root cause in 5 lines"

d) paste the logs into CLAUDE.md

c — the noise stays in the subagent. A strict return format keeps the summary small.

Quiz Progress

0 / 0 questions answered (0%)

0 correct

Quiz Complete!

0%

Reset quiz

Quiz Progress

• Answered: 0 / 0 (0%)
• Correct: 0 / 0 (0%) Reset