# Quiz — Session 6: Claude Code & CI/CD **Program:** Applied AI — Advanced Level — Instructor: Yann Isola **Format:** 10 MCQs (multiple choice questionnaire) — only one correct answer per question — recommended duration: 12 minutes. **Level:** certification depth *Claude Certified Architect* — several questions are judging scenarios. --- **Q1. What fundamentally distinguishes Claude Code from a chat assistant that talks about code?** - A. It uses a larger model, specialized in programming - B. It is an agent: it loops — reads files, modifies, executes commands, observes the results and corrects — within a perimeter of permissions defined by the user - C. It works entirely offline, without calling the model - D. It can only work on one file at a time **Q2. What is the role of the `CLAUDE.md` file?** - A. This is the log of past conversations with Claude Code - B. This is the tool permissions configuration file - C. This is a project instruction file loaded automatically at the start of each session: overview, conventions, common commands, known pitfalls - D. This is the file where Claude Code writes its plans before executing them **Q3. The command `/init` of Claude Code…** - A. Resets all permissions to their default values - B. Inspects the repository and generates a first draft of `CLAUDE.md`, which the team must then rework - C. Initializes a new empty Git repository - D. Clears the context of the current conversation **Q4. In `.claude/settings.json`, the entry `"Bash(npm test:*)"` in the list `allow` means:** - A. All shell commands are allowed if npm is installed - B. Orders starting with `npm test` are allowed without confirmation; other orders `Bash` remain subject to the normal regime (confirmation, or headless refusal) - C. The command `npm test` is prohibited - D. Claude Code will launch `npm test` automatically after each change **Q5. Scenario: a CI pipeline uses Claude Code in headless mode for PR review (reading the diff, producing a comment). What set of permissions should an architect use?** - A. All authorized tools, to avoid blockages in the middle of the night - B. `Read`, `Glob`, `Grep`, `LS`, plus variants `Bash(git diff:*)`/`Bash(git log:*)` — and nothing else: no writing, no `WebFetch`, no free shell - C. Only `Write`, to write the review comment - D. `Bash` without restriction, because the CI is already an isolated environment **Q6. What is the difference in nature between a rule written in `CLAUDE.md` and a hook `PreToolCall`?** - A. None: both guarantee the same level of control - B. The `CLAUDE.md` is more reliable because it is read every session - C. The prompt instruction is probabilistic (the model almost always follows it); the hook is deterministic (code that runs every time and can block the call) — security requirements go into hooks or permissions - D. Hooks only work in headless mode, `CLAUDE.md` only in interactive mode **Q7. The team wants every file modified by Claude Code to be immediately reformatted by the in-house formatter, without exception. What mechanism?** - A. A bold statement in `CLAUDE.md` - B. A hook `PreToolCall` which blocks all edits - C.A filtered `PostToolCall` hook on editing tools (`Edit`, `MultiEdit`, `Write`), which launches the formatter after each modification - D. A hook `Stop` which reformats the entire repository at the end of the session **Q8. In headless mode (`claude -p "..."`), what happens when Claude Code wants to use a tool not authorized by the configuration?** - A. A confirmation window is displayed on the CI runner - B. The tool is automatically authorized after 30 seconds without response - C. The action is refused — there is no one to confirm — and the task may remain incomplete: hence the need to pre-calibrate the permissions to what is strictly necessary for the task - D. Claude Code automatically switches to interactive mode **Q9. Scenario: a test generation CI job has `Write` and runs on a runner containing deployment secrets. The analyzed PR comes from an external contributor. What is the main risk and its priority mitigation?** - A. No risk: the model always refuses malicious instructions - B. Prompt injection: the PR code may contain hidden instructions that the agent will read; mitigation: minimal permissions (limit `Write`, no network tools, no free shell), safeguard hooks, and isolation of the runner from secrets - C. API quota overrun, which is resolved by increasing the timeout - D. The slowness of the model, which we solve with a faster model **Q10. What are `/compact` and `/cost` used respectively in a long interactive session?** - A. `/compact` reduces the size of project files; `/cost` displays the subscription price - B. `/compact` compresses conversation history (summary) to free context and maintain quality; `/cost` displays the consumption of tokens and the cost of the session — two reflexes for controlling long sessions, supplemented by prompt caching which charges stable prefixes at a reduced rate ⚠ (rate: see today's price list) - C. `/compact` merge Git commits; `/cost` estimates development time saved - D. Both commands are only available in headless mode --- ## Fixed and comments | # | Answer | Certification comment | |---|---|---| | Q1 | **B** | The keyword is *agentic*: action loop → observation → correction, under permissions. | | Q2 | **C** | `CLAUDE.md` = permanent and versioned context memory of the project. Canonical structure: overview, conventions, commands, gotchas. | | Q3 | **B** | `/init` produces a *draft*. On the exam, be wary of answers that present automatic generation as a final deliverable. | | Q4 | **B** | Self-acceptance reasons are by order prefix. The `deny` always wins over the `allow`. | | Q5 | **B** | Least privilege: a magazine reads, it does not write. Each superfluous tool widens the attack surface (especially in the face of prompt injection). | | Q6 | **C** | THE session architect distinction: persuasion (prompt) vs. capacity (permissions) vs. control (hooks). | | Q7 | **C** | “After each change” → `PostToolCall` filtered. `Stop` (D) is too late and does not cover intermediate checks. | | Q8 | **C** | Direct consequence of the absence of humans in the loop. Almost certain question under consideration in one form or another. | | Q9 | **B** | Third-party content read by the agent is an injection vector.The defense is not trust in the model but reduction of the perimeter: permissions, hooks, isolation of secrets. | | Q10 | **B** | Management of long sessions: context (`/compact`), cost visibility (`/cost`), structural economy (prompt caching). | **Certification preparation threshold: 8/10.** Below, rework questions Q5, Q6, Q8, Q9 (architect's judgment) as a priority.