Level 2
Answer concrete questions about tool use and web with Workers. Choose one example and explain your design and reasons out loud.
Who can apply
Open to applicants aged 9 and over.
Screen-sharing interview
Share your entire screen through SONUB video, then carry out the tasks.
Level 2: Web is done only with Workers. You are asked how to create a Worker with , check it on your own computer and deploy it.
Common to every level
Example interview question: an AI that guesses the next word
Examiner Build an that guesses the next word while a student types study notes. Explain, step by step, how you would make it and how you would check that it works.
Answer in these four steps. Each step shows what to say, with a short example.
-
1. Goal
Say what the AI does and for whom — what goes in and what comes out. Example: “When a student types ‘machine’, it suggests ‘learning’.”
-
2. Data
Say what it learns from, how much, and what you keep aside for testing. Example: “2,000 sentences from my own notes — 80% to learn from, 20% kept for testing.”
-
3. How it works
Say how it works inside and what you build it with. Example: “A counts which word comes after which and suggests the most frequent next word.”
-
4. Check
Say how you know it works: one number you measure and one limit. Example: “On the test sentences, I count how often the real next word is among the top three suggestions. It looks only one word back, so long sentences go wrong.”
The examiner may change the topic — for example, a filter for spam messages or a lunch-menu recommender — but you answer in the same four steps. Higher levels explain each step in more detail, and the interview also has other questions and tasks for each level.
What to learn
- Choose tools such as , , , or for a task.
- Practise generating video, music and voices with suitable tools.
- Operate tools on the shared screen and explain inputs, settings and results.
How you are marked
- Was the tool appropriate for the task?
- Could the candidate complete the requested themselves?
- Could they evaluate the result and explain concrete settings?
Practice interview questions
These 17 questions are examples of what the examiner may ask in the interview. Do not try to hit a single correct word — explain your answer out loud. Open each one to see a sample explanation and a follow-up question. They are not official exam questions.
Instruction files and /goal
-
1
Examiner Explain how the behaves when there is an AGENTS.md in the , a CLAUDE.md in a parent folder and a CLAUDE.md (~/.claude/CLAUDE.md).
Show a sample explanation
Sample explanation Claude Code reads ~/.claude/CLAUDE.md and every CLAUDE.md it finds walking up from the folder where it started, and joins them from the broadest scope to the most specific (they do not override each other). It reads AGENTS.md directly from v2.1.277, but by default only when there is no CLAUDE.md at all in the working folder or above it (~/.claude/CLAUDE.md does not count for this). Here a parent folder has a CLAUDE.md, so only the global and parent CLAUDE.md files are read and the AGENTS.md in the current folder is not. To read it too, put a one-line CLAUDE.md containing @AGENTS.md in the current folder or change Project instructions in /config, then check with / what was loaded.
Follow-up question How would you set up instruction files so that Claude Code and read the same rules?
-
2
Examiner Explain how to write the (the goal) when you use the command in .
Show a sample explanation
Sample explanation After /goal, write a you can check, not a vague wish such as "make it better". For example: /goal all tests in tests/ pass and the sign-up button calls signup(). Do not change the . Also say what must not change and set limits such as time or , and before you start. Codex marks a goal complete only when evidence such as passing tests supports it, so a goal you can prove works best. While it runs, use /goal to see its status and /goal pause, resume and clear to pause, continue and remove it.
Follow-up question Turn "clean up the " into a goal that suits /goal.
Agents and harnesses
-
3
Examiner Explain why is an , and where and how the official site explains it.
Show a sample explanation
Sample explanation In the official docs (code.claude.com/docs), the "The loop" section of the "How Claude Code works" page explains that when Claude Code takes a task it repeats gathering , taking action and verifying results until the task is done. The Tools section of the same page and the Agentic entry in the Glossary say that tools are what make Claude Code agentic. Without tools a can only answer in text, but Claude Code uses tools to read and edit files and run commands, then looks at the results and decides its next step itself.
Follow-up question Give an example of how this differs from asking about code in a chat window.
-
4
Examiner Explain why is a , and where and how the official site explains it.
Show a sample explanation
Sample explanation The harness entry in the official Glossary defines a harness as the tools, and execution environment that turn a into a capable , and says Claude Code is the harness and is the inside it. The harness supplies file access, execution, gating, and the loop that chains actions together. The "How Claude Code works" page also describes Claude Code as the layer around the model that provides the tools and manages the the model sees, and calls this layer the agentic harness.
Follow-up question What different effects do changing the model and changing the harness have on the results?
-
5
Examiner Explain why is a , and where and how the official site explains it.
Show a sample explanation
Sample explanation In the official blog post "Unrolling the Codex loop" (openai.com), explains that the Codex harness provides the core agent loop and execution that every Codex experience is built on. The harness runs the cycle: user input → → either a response that ends the turn, or whose results go back for another round of inference. So even a short greeting goes through this loop. The "" post on the same site calls the work of designing the environment, instructions and — so that people steer and Codex executes — harness engineering.
Follow-up question Why can you say that even sending Codex the single word "hello" goes through the harness?
Skills and settings files
-
6
Examiner Explain how a (SKILL) differs from a skill under agentskills.io.
Show a sample explanation
Sample explanation created and published the skill format, and it is now kept as the at agentskills.io. The standard only defines what goes into a folder with a , and its standard frontmatter fields are six: name, description, , compatibility, and . Claude Code follows the standard but adds fields such as disable-model-invocation, : fork and paths. It also looks for skills in (project) and ~/.claude/skills/ (personal) rather than the ./skills/ the standard recommends, and when names clash a personal skill wins over a project skill. For skills shared with other tools, use only the six standard fields.
Follow-up question How do you set up a skill in Claude Code so that it runs only when you call it with a /command and the never uses it on its own?
-
7
Examiner Explain where the settings files of and are, and how to set the and the auto-compact size in them.
Show a sample explanation
Sample explanation Claude Code reads ~/.claude/settings.json (all your projects), .claude/settings.json (the team's committed file) and .claude/settings.local.json (just you, in this project); a higher level wins in the order > project > user. Set the model like "model": "opus" and the auto-compact size as a count like "autoCompactWindow": 444000 (100000 to 1000000, or "auto"). Codex uses ~/.codex/config.toml (project settings in .codex/config.toml) with model = "<model name>" and model_auto_compact_token_limit = 475000. After changing them, check with /status that they applied.
Follow-up question What goes wrong if you set the auto-compact size too large, and what if you set it too small?
Comparing tools, and security
-
8
Examiner Explain what Projects are and how they differ from .
Show a sample explanation
Sample explanation A Claude project is a workspace for one goal: it keeps instructions, knowledge files and chats together so any chat starts from material you already gave. Classic projects live in Claude chat (claude.ai), where you have one chat at a time; a new chat starts from the project’s instructions and knowledge, and reaches earlier chats only through project or past-chat search, when your plan and settings turn them on. Claude Code is a that reads and edits files and runs commands directly in a project folder in your . The new Projects (beta) run inside Claude Code: with instructions, memory and , Claude starts several threads at once, and they are rolling out to some Pro and Max users.
Follow-up question What must you never put in a project's knowledge files, and why?
-
9
Examiner Explain how Duet and Dot differ.
Show a sample explanation
Sample explanation Claude Duet is 's app that comes to you first: it sends a summary every morning, works on its own computer through the apps you connect for everyday tasks such as mail and calendars, and always asks before sending, paying, signing or deleting. Codex Dot (a dot) is 's always-on personal agent inside : with its own cloud computer and it keeps working after the chat ends, and it hands work to Codex tasks (cloud, or an existing thread), steering and checking them. Both act within the rules and approvals people set, so a person must read the drafts and approval cards.
Follow-up question Why should you tell an agent "write a draft" and "send it" as separate requests?
-
10
Examiner Explain how you would set up when you hand work to an tool.
Show a sample explanation
Sample explanation Start with narrow : have it ask before editing files and running commands, and allow only safe commands such as tests. Block secret files: in add "Read(./)" to permissions.deny in settings.json, and in use sandbox_mode and in config.toml to set which folders and it can use and when it must ask. Keep passwords, and personal out of the work folder, and add only you trust. Watch for instructions hidden in outside text (), and have a person review results before they leave the team.
Follow-up question If you accidentally showed an API key to the AI or pushed it to a , what would you do first?
Speech recognition (STT) and speech synthesis (TTS)
-
11
Examiner Explain what (speech recognition) is and how exactly you can use it at work.
Show a sample explanation
Sample explanation STT (speech-to-text) turns spoken words into text. Use it to turn recordings of meetings, classes and consultations into minutes and summaries, to add subtitles to videos, and to dictate notes and commands when your hands are busy. To use it at work, get consent before recording, record with little noise, and have a person check and fix parts that are easy to get wrong, such as technical terms and names. For recordings with personal , check which they are sent to, and if needed use a that runs on your own computer (for example, Whisper).
Follow-up question How would you quickly find and fix wrong words in the transcript?
-
12
Examiner Explain what (speech synthesis) is, and exactly what you must do, and how, to use it at work.
Show a sample explanation
Sample explanation TTS (text-to-speech) reads text aloud in a human voice. Use it for narration in lecture and guide videos, customer guidance messages, and reading text aloud for people who find reading hard. First write the script, choose the language, voice and speed, rewrite numbers, abbreviations and names the way they should be read, then generate it and listen to the end to fix wrong pronunciation. Check the plan and terms of use to see whether commercial use is allowed, say that the voice is AI-made, and clone someone else's voice only with that person's clear consent.
Follow-up question What would you check first when making the same guidance message in several languages?
Deploying with Cloudflare Workers
-
13
Examiner Explain what Workers is and why level 2 websites only with Workers.
Show a sample explanation
Sample explanation Workers is a serverless that runs your on Cloudflare's around the world. You do not rent or manage a server, and small sites can start within the free tier. Static files such as , , and images go up together with the code that handles requests as one Worker, so you only need to learn one way to deploy properly. Prices and free tiers change, so check the official pricing page.
Follow-up question Where would you check what happens when you go over the free tier?
-
14
Examiner Explain the order of commands to create a new Worker project, check it on your own computer and it.
Show a sample explanation
Sample explanation Create the project with create cloudflare, cd into its folder, and check it first on your computer with npx dev. Log in to your account with npx wrangler login and publish with npx wrangler deploy; you get an address such as name.account.workers.dev.
Follow-up question If the page has not changed after deploying, what would you check first?
-
15
Examiner Explain what goes into the (wrangler.jsonc or wrangler.toml).
Show a sample explanation
Sample explanation It holds the Worker's name (name), the file to run first (main), the date (compatibility_date), the static files folder ( under assets) and connections to storage such as KV and D1 (bindings). This file is the blueprint of the , so if an changed it, a person reads it and checks what changed.
Follow-up question What can change when you change compatibility_date?
-
16
Examiner Explain how to use secret values such as safely in a Worker.
Show a sample explanation
Sample explanation Do not write secrets in or that go into the ; put them into separately with npx secret put <NAME>. In code, read them as env.<NAME>. For checks on your own computer keep them in a .dev.vars file, and add that file to .
Follow-up question If a was pushed to by mistake, what would you do first?
-
17
Examiner Explain how to connect your own to a Worker and how to when a goes wrong.
Show a sample explanation
Sample explanation Once the domain is added to , connect it in the Worker's settings on the or with routes (custom_domain) in the . If a deployment goes wrong, list past deployments with npx wrangler deployments list and go back to the previous with npx wrangler rollback. Watch live errors with npx wrangler tail.
Follow-up question After rolling back, which records would you look at to find the cause?