Level 4 — Explore, Plan, Implement, Commit

10 questions, drawn from a bank of 24 across 21 topics. Pass at 80%. 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 / 10 · Choose one
/diff shows a change to a config file you are certain Claude never touched this session. Is the diff wrong?
/diff shows a change to a config file you are certain Claude never touched this session. Is the diff wrong?
No. /diff shows the edits Claude has made so far alongside anything else you have not committed — including the experiment you made by hand and forgot aboutNo, but it means Claude probably did touch it — through a shell command, the kind of change that would only ever have shown up under CurrentYes — you are reading a checkpoint snapshot of the files as they stood before your last prompt, not the working tree as it is nowYes — /diff is meant to show this session's edits, so a file Claude never opened appearing in it means the view is out of date
2 / 10 · Choose one
You read a stack trace, conclude the fix is a missing null check, and type "add a null check in user.ts". Claude does it correctly and fast. What have the docs warned you about?
You read a stack trace, conclude the fix is a missing null check, and type "add a null check in user.ts". Claude does it correctly and fast. What have the docs warned you about?
That a request whose verb is an edit is always wrong, and every fix belongs in plan mode firstNothing — the gather-context phase runs on every task, so Claude will search the surrounding code and find the real cause before it writes the null check you asked forLetting Claude jump straight to coding can produce code that solves the wrong problem — a correct implementation of your own diagnosisNothing much — skipping exploration trades a little accuracy for speed, and on a fix as small as a single null check that is the right trade to make
3 / 10 · Choose one
Two counts get confused in this level: the agentic loop has three phases and the recommended workflow has four. Which trio is the loop?
Two counts get confused in this level: the agentic loop has three phases and the recommended workflow has four. Which trio is the loop?
Gather context, take action, verify results — search, edit, then check the workExplore, plan, implement — read the code, write the plan, then make the edits it namesPlan, act, observe — decide the next step, call one tool, then read what the tool returnedRead, edit, test — open the file, apply the diff, then run the suite to confirm nothing broke
4 / 10 · Choose one
The plan on screen names the wrong file in step two. You know exactly what it should say. What does Ctrl+G give you that typing the correction into chat does not?
The plan on screen names the wrong file in step two. You know exactly what it should say. What does Ctrl+G give you that typing the correction into chat does not?
It sends your correction to Claude as a structured plan-revision instruction, which the implementing turn follows more reliably than free textIt writes the plan to ~/.claude/plans so you can open the file and edit it after the session rather than at the promptIt opens the proposed plan in your default text editor so you edit it directly before Claude proceeds — the corrected plan becomes the artifact rather than an amendment sitting above it in the conversationIt makes Claude regenerate the plan from scratch with your change applied, so you approve a fresh proposal rather than a hand-patched one
5 / 10 · Choose one
A session has showClearContextOnPlanAccept turned on. During planning you typed "leave the legacy exporter alone" into the chat, never adding it to the plan text. You approve with the first menu option. What reaches the implementation?
A session has showClearContextOnPlanAccept turned on. During planning you typed "leave the legacy exporter alone" into the chat, never adding it to the plan text. You approve with the first menu option. What reaches the implementation?
Everything — the option clears the conversation only once the implementation finishes, so it is the *next* task that starts in a window this one has not spentThe plan alone — that option approves the plan, clears the conversation context, and starts implementing from the plan aloneThe plan plus your stated constraints — the clear drops the reasoning and the tool output, and carries the instructions you gave during planning into the implementationThe plan alone — and because showClearContextOnPlanAccept is on by default, every session that approves with the first menu option behaves this way
6 / 10 · Choose one
You are about to send: "The checkout flow is broken for users with expired cards. The relevant code is in src/payments/. Can you investigate and fix it?" A colleague says you should list the files Claude ought to read. What do the docs say?
You are about to send: "The checkout flow is broken for users with expired cards. The relevant code is in src/payments/. Can you investigate and fix it?" A colleague says you should list the files Claude ought to read. What do the docs say?
Rewrite it — "investigate and fix it" is the infinite-exploration antipattern, whose documented fixes are to scope it or hand it to a subagentSend it as written — you do not need to specify which files to read or what commands to run, because Claude figures that out from a symptom and a locationAdd the file list — naming the two or three files under src/payments/ that handle expiry is strictly cheaper, because Claude reads exactly those and runs no search at allAdd the file list — Claude works from the file you have open, so it cannot search the repository for the ones you did not open
7 / 10 · Choose one
You type "switch to plan mode and don't touch any files". Claude replies that it is now planning. What tells you whether that is true?
You type "switch to plan mode and don't touch any files". Claude replies that it is now planning. What tells you whether that is true?
Watch the next few turns — plan mode is the thing that stops an edit, so if Claude reads files and proposes changes without writing any, the mode is on and the request workedNothing in the session shows it; you would have to quit and relaunch with claude --permission-mode plan to be certainThe status bar. Asking Claude in chat to change the permission mode does not work, so the only evidence is the indicator reading ⏸ plan mode onThe reply itself — Claude read the instruction and agreed to it, and an instruction Claude agreed to binds the session exactly as a mode set with Shift+Tab does
8 / 10 · Choose one
A teammate refuses the clear-context approve option: "clearing wipes the plan too, and then Claude is implementing from memory." What is the accurate correction?
A teammate refuses the clear-context approve option: "clearing wipes the plan too, and then Claude is implementing from memory." What is the accurate correction?
Backwards — nothing is lost either way, since a cleared conversation is summarised into the new window the way /compact summarises a long session, so the planning discussion rides alongThey are right, and the safe habit is to take "Yes, manually approve edits" instead, so the planning conversation stays available to the implementationThey are half right — the plan is cleared with everything else, but Claude re-injects it from disk on the next turn the way it does after a compactionBackwards — the plan is what survives, because that option implements from the plan alone. It buys a window planning has not already spent and costs the conversation around the plan
9 / 10 · Choose one
You have been asked to rename a config key across roughly thirty files. One documented planning signal is present — the change modifies multiple files. Do you plan?
You have been asked to rename a config key across roughly thirty files. One documented planning signal is present — the change modifies multiple files. Do you plan?
No. The docs put the one-sentence-diff test in the same passage as the signals, and every changed line here is the same change — a plan would be thirty lines saying one thing thirty timesYes. Planning is cheap insurance and a plan you turn out not to need costs only the minute it takes to read it, so on anything touching thirty files it should be the defaultYes. The docs' own examples of work to ask for directly are a typo and a log line — single-file edits — and thirty files is far past thatYes. "The change modifies multiple files" is one of the three documented signals, and the docs join the signals with *or*, so one of them being present settles it
10 / 10 · Choose one
You delegated an import cleanup to a background subagent, disliked the result, and rewound the session. What does the documentation say happened to those edits?
You delegated an import cleanup to a background subagent, disliked the result, and rewound the session. What does the documentation say happened to those edits?
They are restored like any other edit, because a subagent makes its edits with Claude's file editing tools and checkpointing tracks every change those tools makeThey are restored as long as the delegation falls inside the 100 most recent checkpoints in the session, which one import cleanup comfortably doesNothing at all is restored — a rewind across a subagent run fails outright with No files were restored, leaving the tree as it wasRewinding does not restore them — for any subagent but a foreground forked skill, use git to revert the edits

10 questions still to answer