Level 3 — Context Is the Budget

12 questions, drawn from a bank of 34 across 25 topics. Pass at 85%. Every wrong option carries a note explaining why it is tempting — those notes are the point of the quiz, so read them whether you pass or not.

attempt 1

1 / 12 · Choose one
You have corrected the same issue three times, and you run /clear. What is the other half of the documented instruction — the part most people skip?
You have corrected the same issue three times, and you run /clear. What is the other half of the documented instruction — the part most people skip?
Compact first with instructions naming the bug, so the fresh window inherits a summary of what has already been ruled outRe-send the original prompt unchanged into the clean window, so the agent gets an uncontaminated attempt at exactly what you asked forStart fresh with a more specific prompt that incorporates what the failed corrections taught youWrite the correction into the project CLAUDE.md, where it is re-injected from disk and holds for every future session
2 / 12 · Choose one
You opened a session at 9am and have never filled it enough to compact. At 5pm, after a lunch break, you type one word: "why?". What does that request actually carry?
You opened a session at 9am and have never filled it enough to compact. At 5pm, after a lunch break, you type one word: "why?". What does that request actually carry?
Only the messages since the last automatic compaction, since everything before it was replaced by the summaryOne word of input, plus whatever the answer costs — the morning's turns were processed hours ago and do not need sending a second timeThe whole day's conversation, re-sent — and because an idle hour has expired the prompt cache, reprocessed at full priceThe whole conversation, but at the cached rate, which holds as long as the session stays open on a subscription
3 / 12 · Choose one
Which situation is /btw designed for?
Which situation is /btw designed for?
Starting an unrelated task without polluting the current one, since the side thread gets a context of its own to work inRouting a cheap lookup to a small, fast model so a one-line question does not spend Opus tokensA detail you need once — "what version are we pinned to?" — whose answer you do not want sitting in the window for the rest of the sessionAsking something you would rather keep off the record — the exchange happens outside the session, so it leaves no trace and no usage
4 / 12 · Choose one
Your input box has half a prompt typed into it. You press Esc Esc expecting the rewind menu. What happens?
Your input box has half a prompt typed into it. You press Esc Esc expecting the rewind menu. What happens?
The draft is submitted — with text in the box, the second Esc acts as EnterThe rewind menu opens and your draft is preserved underneath it, ready for you when you come back outThe draft is discarded permanently; there is no undo for the input boxThe draft is cleared and saved to input history — press Up to recall it, then Esc Esc again on the now-empty box for the menu
5 / 12 · Choose all that apply
Which of these does checkpointing not restore? Select all that apply.
Which of these does checkpointing not restore? Select all that apply.
Changes made by Bash commands such as rm, mv or cpYour own manual edits, and edits from other concurrent sessionsEdits made through Claude's file-editing tools in the current sessionSymlinked and hard-linked paths, skipped with a Restored the code, but skipped N files warningEdits made by a subagent, other than a foreground forked skill
6 / 12 · Choose one
You type /clear release-prep. Which conversation does that name attach to?
You type /clear release-prep. Which conversation does that name attach to?
The new, empty conversation, exactly as claude -n release-prep would have named itThe one you are leaving — it labels the previous conversation in the /resume pickerBoth conversations, so one label follows the workstream across the clearNeither — it is a note recorded in the session log, not a name in the picker
7 / 12 · Choose one
Claude has just re-suggested an approach you explicitly rejected an hour ago, in a session that has run all afternoon. Which reading does the documentation support?
Claude has just re-suggested an approach you explicitly rejected an hour ago, in a session that has run all afternoon. Which reading does the documentation support?
Performance degrades as the window fills, so the rejection is buried in a window too full to hold it — the response is to clear, not to repeat yourself more firmlyThe model is having a bad day. Nothing about the session has changed since this morning, so the move is to send the correction again and let it pick the thread back upYour original rejection named the approach but never said what to do instead, so it was too terse to generalise and needs restating more specificallyIn-chat constraints never hold for long. A rejection has to go into the project CLAUDE.md, which is re-injected from disk after every compaction, or it will not survive the afternoon
8 / 12 · Choose one
A session was launched with --mcp-config ./mcp.json. You resume it and the MCP tools are missing. Why?
A session was launched with --mcp-config ./mcp.json. You resume it and the MCP tools are missing. Why?
It is a bug — resuming restores the conversation, the model, the agent, the permission mode, an active goal and unexpired scheduled tasks, so launch flags should come back tooYour settings.json was re-read at launch and overrode the flagResuming does not restore that flag — --mcp-config, --settings, --plugin-dir, --fallback-model and --add-dir all have to be passed againMCP server connections are torn down with the session, and a resumed conversation cannot re-establish them
9 / 12 · Choose one
After a compaction, which of the files Claude had read come back into context?
After a compaction, which of the files Claude had read come back into context?
Up to five, most recently modified first — and one over about 5,000 tokens returns as a path reference rather than its contentsNone. A file comes back only when Claude reads it again, which is why the first turn after a compaction usually repeats the same readsAll of them — re-reading a file from disk is cheaper than summarising it, so nothing Claude has already read needs to be given upAll of them, as a one-paragraph summary per file, so the shape of everything read survives compaction at a fraction of the token cost
10 / 12 · Choose one
You want yesterday afternoon’s work back, in a directory where you have run four other sessions since. Which command?
You want yesterday afternoon’s work back, in a directory where you have run four other sessions since. Which command?
Run claude, then /clear yesterday to bring that conversation backThere is no way back — transcripts are not searchable by day, so start fresh and re-explainclaude --resume <name> if you named it — otherwise claude --resume and pick it out of the listclaude --continue, which resumes the most recent session in this directory
11 / 12 · Choose one
What is /context documented as doing?
What is /context documented as doing?
Listing every file the session has read, with each one's token cost, so you can decide which to re-read after a compactionShowing session cost, plan usage limits and activity stats, with a breakdown of the behaviours counting most against your plan limits, such as long contextVisualising current context usage as a coloured grid, with optimisation suggestions for context-heavy tools, memory bloat and capacity warningsCompacting the conversation once usage passes the threshold, then reporting how much room the summary freed
12 / 12 · Choose one
Which of these describes the documented "kitchen sink session" failure pattern?
Which of these describes the documented "kitchen sink session" failure pattern?
Starting one task, asking something unrelated, then going back to the first — leaving context full of irrelevant informationAny session left open for more than about two hours, by which point the earliest turns no longer describe what you are working onRunning more MCP servers than the task requires, so their tool definitions crowd the window from startup onwardAsking the agent to investigate something without scoping it, so it reads hundreds of files and fills the window

12 questions still to answer