Articles
Codility Claude Code Interview Guide 2026: Shared Sessions, Prompt Replay, and Candidate Reports
Prepare for a Codility Claude Code interview: shared terminal sessions, prompt replay, employer reports, verification, and a seven-day practice plan.
A Codility Claude Code interview is an AI-assisted coding session in which an employer can enable Claude Code inside Codility's hosted terminal. In a live Interview, the candidate and interviewer can share one synchronized session; prompts, responses, and resulting activity can later appear in Replay and the employer-facing candidate report.
The feature is not guaranteed for every Codility link. Account administrators and interview creators control whether AI assistance is available, and the August 2026 release is still described as Preview. Treat your invitation and the instructions shown before the session as the source of truth.
For scoring and hidden tests, use the separate Codility Assessment Guide 2026.
Quick answer: what happens in a Codility Claude Code interview?
Codility's August 2026 product update says Claude Code can run inside Screen, Interview, and Skills Intelligence when the employer enables it. In Interview, both participants can see a shared Claude Code conversation. In later review, Codility can preserve the AI exchange, session activity, and a per-task usage summary.
| Feature | What Codility currently documents | What you should do |
|---|---|---|
| Employer-controlled access | AI tools can be enabled or disabled for an assessment or interview | Never assume an external AI tool is allowed when the built-in option is absent |
| Shared live session | Candidate and interviewer see the same Claude Code prompts and responses in Interview | Keep prompts specific and narrate corrections rather than hiding uncertainty |
| Hosted setup | Claude Code is preinstalled and preconfigured in the Codility environment | Do not spend interview time on personal credentials or local setup |
| Replay | AI interactions and session activity can be reviewed after the round | Make requirements, tests, and decisions visible in the work itself |
| Candidate report | Reviewers can see AI usage summaries and open the full transcript | Optimize for engineering judgment, not a low prompt count |
The company still chooses the repository, task, timer, rubric, allowed resources, and whether Claude Code is enabled.
What changed in Codility's August 2026 release?
The most important change is that Claude Code moved from an Interview-only early preview toward a broader Preview across Screen, Interview, and Skills Intelligence. Codility says the tools are off by default at the account level; administrators and assessment creators control access.
For each task where AI was used, the report can summarize the model, interaction count, and token count before a reviewer opens the transcript.
Some older Codility pages still describe an Interview-only roadmap. Prefer the dated August 4 release note, while remembering that availability varies by employer and account.
How the shared Claude Code session works
In a live Codility Interview, Claude Code runs in the same hosted VS Code environment as the task. Codility's shared-session product page says the candidate can open the terminal and run claude; the assistant is already configured, so there is no personal API key, separate login, or onboarding dialog.
The interviewer and candidate share one synchronized AI session. Both can see prompts and responses, making the prompt part of the technical conversation rather than a private shortcut.
A strong candidate uses that visibility deliberately:
- Clarify the behavior before asking the agent to edit.
- Inspect the smallest relevant repository area and state uncertain assumptions.
- Request one bounded change rather than a broad rewrite.
- Review the diff, run independent tests, and explain what remains unverified.
If the agent proposes the wrong approach, say so and redirect it. A visible correction is positive evidence that you are controlling the tool instead of accepting its first answer.
Screen, Interview, and Skills Intelligence are not identical
Codility documents Claude Code across three product surfaces, but candidates should not flatten them into one experience.
| Surface | Session shape | Likely review context |
|---|---|---|
| Screen | Employer-configured assessment completed by the candidate | Automated task results plus saved AI activity and report evidence |
| Interview | Live shared VS Code and Claude Code session with interviewer participation | Real-time reasoning, collaboration, code, tests, and later Replay |
| Skills Intelligence | Skills assessment for employees or candidates | Skills evidence plus the same AI-usage reporting layer where enabled |
The official release does not promise the same task type, duration, or scoring model across these surfaces. A Screen is less conversational than a live Interview, where you should expect to explain decisions while the other participant watches.
What prompt replay and the candidate report can show
Prompt replay is the reviewable record of how AI assistance affected the work, not merely a copy of the final code. Codility says the saved evidence includes what the candidate asked, what Claude Code answered, and what the candidate did with the answer. Replay lets a reviewer move through that sequence.
The employer-facing candidate report can add a per-task AI usage summary. Codility currently documents three fields: the model used, the number of interactions, and token usage. The complete transcript remains available behind the summary.
Do not over-interpret these fields. Codility has not published a rule that fewer prompts are better, a token threshold is suspicious, or one interaction count decides the outcome. Transcript quality depends on the investigation, corrections, and testing it documents.
The candidate report is employer-facing; public documentation does not establish that every candidate receives the transcript or report.
What reviewers are likely evaluating
Codility frames the signal as how a candidate prompts, corrects, and turns an initial result into working code. The following is practical inference, not an official scoring formula.
Requirement control. Identify the contract, constraints, and success conditions before delegating. Strong prompts state behavior and boundaries instead of asking the model to “solve everything.”
Repository judgment. Inspect the architecture, dependencies, and conventions. Finding the right change surface matters more than generating a large diff.
Verification. Challenge output with independently chosen tests. Anthropic's Claude Code security guidance says users remain responsible for reviewing code and commands; make that responsibility visible.
Correction quality. Recognize invented APIs, weakened assertions, unsafe commands, or missed edge cases and recover cleanly.
Ownership and communication. Explain the final code without reading the transcript and defend every important line, test, and trade-off.
A reliable prompt-to-proof workflow
Start with a read-only exploration prompt:
Trace the request through the repository. Name the relevant files, existing tests, and uncertain assumptions. Propose the smallest safe change and do not edit yet.
Check the response against the repository. If it missed an authorization layer, data migration, generated file, or integration boundary, correct the scope explicitly. Then ask for one change with an objective test:
Implement only the validated change in these files. Preserve the public interface. Add a failing test for the reported behavior first, then make it pass. Do not alter unrelated formatting or dependencies.
After the edit, inspect the diff yourself. Run the narrow test, then the relevant suite, lint, or type check when time allows. Add one adversarial case the model did not suggest. Finish by summarizing the behavior, changed files, evidence, and residual risk in your own words.
This workflow mirrors the Claude Code agentic loop: gather context, take action, verify results, and iterate. The interview skill is deciding when each phase has enough evidence to move forward.
Common mistakes in an AI-assisted Codility round
First, Claude Code support does not permit every AI tool. If the built-in assistant is disabled, follow the displayed rules and ask before using anything external.
Second, do not start with implementation. Without repository context, the agent may invent paths, duplicate abstractions, or optimize the wrong behavior. Inspect and plan first.
The third is confusing generated tests with independent proof. Derive at least one case yourself from the contract, boundaries, or a failure path.
Fourth, do not hide a bad turn. A concise correction is better: “That changes the prohibited public API. Revert that part and keep the existing signature.”
Meaningful prompts, bounded edits, and verification matter more than an artificially short transcript.
A seven-day preparation plan
Day 1: Review your invitation, the general Codility interface, and the exact AI policy. Practice starting from requirements rather than a prompt.
Day 2: Map files, dependencies, and execution in an unfamiliar repository without editing.
Day 3: Use Claude Code for one bounded bug fix. Require a test first and reject any unrelated refactor.
Day 4: Review an AI-generated diff for contract changes, error handling, security, and weak assertions.
Day 5: Run a timed feature task and narrate assumptions, corrections, tests, and residual risk.
Day 6: Review your transcript for vague prompts, repetition, and output accepted without evidence.
Day 7: Complete a full mock and deliver a two-minute defense of the implementation.
Frequently asked questions
Does every Codility interview allow Claude Code?
No. Codility supports Claude Code when the account administrator enables the feature and the assessment or interview is configured to allow it. The August 2026 rollout is still described as Preview, so availability can differ. If Claude Code is absent, do not substitute an external assistant unless the invitation explicitly permits one.
Can the interviewer see my Claude Code prompts live?
In a Claude Code-enabled Codility Interview, Codility says interviewer and candidate share one synchronized session and can see prompts and responses as they occur. A Screen assessment is not the same live collaboration format, but AI activity can still be saved for later report review when the feature is enabled.
What appears in the Codility candidate report?
Codility documents saved Claude Code interactions, full Replay, and a per-task AI usage summary that can include the model, interaction count, and token usage. The full transcript remains available to reviewers. Exact report fields can evolve during Preview, and public documentation does not publish one universal AI score.
Is using fewer prompts better?
Not necessarily. Codility does not publish a rule that rewards the lowest interaction count. Reviewers need enough evidence to understand how you scoped, corrected, and verified the work. Prefer focused prompts with clear constraints, but do not avoid a necessary correction or test merely to make the transcript shorter.
How should I prepare if I have never used Claude Code?
Learn a simple loop: inspect the repository, state a plan, request one bounded change, review the diff, run independent tests, and explain the result. You do not need clever prompt tricks. You need enough tool fluency to keep the session moving while retaining engineering ownership of every accepted change.
Final takeaway
A Codility Claude Code interview makes AI use visible rather than private. In Interview, the shared terminal exposes prompts and corrections in real time; Replay and the candidate report can preserve the same evidence for later review. The employer still controls whether the feature exists and how the session is evaluated.
Prepare for judgment, not prompt theater. Clarify the contract, inspect before editing, constrain the agent, verify independently, and defend the final code.
Sources and Further Reading
- Codility: You can now use Claude Code in Codility — August 2026 update
- Codility Interview: Claude Code in the terminal
- Codility: AI in technical assessment — a working framework for 2026
- Codility Support: AI Interview Assistance Settings
- Anthropic: Claude Code overview
- Anthropic: How Claude Code works
- Anthropic: Claude Code security
Research note: This guide was checked on September 1, 2026. Codility features, Preview availability, employer settings, and interview rules can change; your invitation and in-session instructions are authoritative.