Job search
Anthropic Application AI Policy: What You Can Draft with AI and What Must Be Your Own
Anthropic’s public candidate guidance asks applicants to create their first draft themselves, then use Claude or another AI tool to refine it.
Anthropic’s public candidate guidance asks applicants to create their first draft themselves, then use Claude or another AI tool to refine it. Your experience, perspective and factual claims must remain your own. The policy separately restricts AI during take-home assessments and live interviews unless permission is given, so permission to polish application language does not extend automatically to every stage of hiring.
LandOffer publishes this guide. When you browse recently posted jobs, read each employer’s current assistance rules before working on its application. The policy described here was checked October 12, 2026; Anthropic’s guidance is dated July 10, 2025. The worked drafts below are fictional teaching examples, not application answers written for a real candidate.

Read the current guidance and the actual application together
Read Anthropic’s candidate policy together with the specific role’s application instructions. The public policy distinguishes application refinement, interview preparation, take-home work and live interviews. It discourages invented experience and expects transparency in following the guidance. Check for task-specific instructions before using assistance, especially if a later exercise expressly defines what is allowed. Anthropic candidate AI guidance
A public Staff+ Research Systems Engineering application checked for this guide links the policy and includes a required AI-policy acknowledgment. The observed public form supports only that role-specific acknowledgment; no authenticated submission or universal form layout is claimed. Observed public application
Anthropic’s careers page also directs candidates to its AI guidance. Use the employer’s current materials rather than an older forum summary or a remembered claim that all AI use is prohibited or encouraged. The relevant question is what this employer permits for this stage and task. Anthropic careers
If the task instructions appear to conflict with general guidance, ask the recruiting contact before proceeding with the disputed assistance. Do not resolve the conflict by choosing whichever interpretation allows more automation. Save the instruction and the stage it covers so you do not reuse permission outside its scope.
Build your first draft from facts you can explain
Begin with your own notes about the role and your actual work. Identify a project, your contribution, the decisions you made and a result you can support. Write a first version in your own words before asking a tool to improve it. Use the first draft to record your thinking and actual work before refining the language.
Keep the factual inputs narrow. A repository, performance review or project note may help you remember your contribution, but do not upload confidential employer material into a tool without permission. You can write a permitted factual summary that omits sensitive details while still distinguishing your role from the team’s work.
Separate facts from interpretation. “I implemented the retry handler” is a contribution. “The team’s deployment became more reliable” is a broader result that needs a basis. “This made me interested in reliability work” is a personal reflection you should be able to explain. All three can belong in an answer, but they should not be blended into an unsupported claim of sole ownership.
If you cannot identify a measured result, describe a supported outcome or limitation. A truthful statement that you added a recovery path and documented its behavior is preferable to an invented percentage improvement. A plausible generated number still needs evidence from your actual work.
A completed first-draft and refinement example
This fictional teaching case concerns Nina, an engineer considering a systems role. Her permitted factual notes say she implemented a retry handler for one service, added integration tests and wrote a runbook. A teammate designed the overall migration. No verified percentage improvement is available.
Nina writes her first draft herself:
On our service migration, I implemented the retry handler, added integration tests for transient failures and wrote the recovery runbook. Another engineer designed the overall migration. The work made me more interested in how recovery behavior is tested and explained to operators.
She then asks for a bounded edit: “Review my draft for clarity. Keep the contribution and limitations unchanged. Do not add metrics, tools, leadership claims or experiences. Identify any sentence that needs evidence rather than filling the gap.”
| Element | Supported input | Accepted refinement |
|---|---|---|
| Ownership | Retry handler for one service | “I implemented the service’s retry handler” |
| Testing | Integration tests for transient failures | Kept as a specific action |
| Documentation | Recovery runbook | Kept as a specific deliverable |
| Migration design | Teammate’s responsibility | Not reassigned to Nina |
| Result metric | None verified | No percentage added |
Her reviewed version reads:
During a service migration, I implemented the retry handler, added integration tests for transient failures and wrote the recovery runbook. A teammate owned the migration design. My contribution focused on making recovery behavior testable and understandable to the engineers operating the service.

The revision makes Nina’s recovery work easier to understand while preserving the limits of her contribution. Nina still needs to decide whether it answers the actual application question and reflects what she wants to communicate. The example is deliberately not a finished “Why Anthropic?” response for a reader to paste into a real form.
Audit every new claim introduced by an edit
Compare the edited version with your original line by line. Look for stronger verbs, expanded scope, added numbers, assumed technologies and personal motivations you did not express. A suggestion can change meaning while appearing to be a stylistic improvement.
Check who did the work. “We shipped” can become “I led” if you accept an edit without review. Check the scale too: a single-service contribution can become an organization-wide initiative. Check whether a proposed outcome is measured, observed or merely hoped for. Each change should have an actual basis before it survives into the application.
Watch for generic personal claims. A sentence saying you have always been committed to a field may not reflect your history. Replace it with a specific truthful connection you can discuss, or remove it. An application should not manufacture a motivation because the employer’s mission sounds attractive.
Do not ask the tool to repair missing facts by inventing likely details. Mark the missing information and return to your records or own recollection. If a fact cannot be supported, narrow the claim. Keeping a limitation can make an answer more accurate and easier to defend in conversation.
A completed rejection of an unsupported AI rewrite
In a second fictional case, Omar writes that he maintained a dashboard used by his immediate team. He asks an editing tool to improve clarity. Its proposed version says he “led an enterprise analytics transformation that increased executive decision speed by 40%.” None of those expanded claims appears in his factual notes.
| Proposed change | Why Omar rejects it | Supported replacement |
|---|---|---|
| Led transformation | He maintained a dashboard, not the overall program | Maintained the team dashboard |
| Enterprise scope | Only his immediate team’s use is established | Used by the operations team |
| Executive users | Not established | Removes the user claim |
| 40% improvement | No measurement exists | Removes the percentage |
| Decision-speed result | Not tracked | Describes the maintained reporting function |
Omar’s final teaching sentence is: “I maintained the operations team’s dashboard, corrected recurring data-quality issues and documented the weekly refresh process.” The last two actions are stipulated facts in this example, not additions inferred from maintaining a dashboard. In a real application, Omar would retain them only if they were true.
His record of the editing session notes which suggestions were rejected and why. This is useful for his own review, but the article does not claim that Anthropic requires every applicant to upload a prompt history or editing log. Follow any actual disclosure request and answer it truthfully rather than inventing a disclosure requirement.
Omar’s completed edit retains dashboard maintenance and removes unsupported enterprise scope and the 40% result. It does not promise that the employer will prefer the wording or advance Omar. Accuracy and policy compliance are necessary review questions; hiring remains a separate decision.
Use preparation without importing it into restricted work
Keep separate files or notes for application editing, interview preparation and any assessment. Label each with the instruction that governs it. This simple separation can prevent a permitted preparation tool from remaining active during a stage where assistance is not allowed.
Before a take-home exercise, read the instructions again. Identify whether AI is expressly permitted, prohibited or not addressed. If you need clarification, ask before using the tool on the task. Do not treat a general statement about collaborating with AI as permission to generate assessment code or answers.
Before a live interview, close tools that would provide unapproved real-time help. Preparation notes are not automatically authorized during the interview itself. If you need an accommodation, discuss it through the employer’s designated process instead of assuming a private workaround is permitted.
A practice explanation can help you discover gaps in your understanding. Use that discovery to study and improve your own reasoning before the evaluated task. Do not present a generated explanation as evidence that you personally solved the problem or performed the underlying work.
Protect confidentiality while seeking useful feedback
Use only information you are allowed to share. Remove private customer data, credentials, unreleased designs and restricted internal material before using a tool. Even a request for grammar help can expose information if the source text contains secrets. A safer summary still needs to remain truthful about your role and the project.
Avoid uploading someone else’s evaluation or recommendation letter merely to strengthen an answer. Consider the document’s confidentiality and the writer’s permission. If a public portfolio is relevant, verify that you can share it and that the link works without granting broad access to private workspaces.
For claims you cannot disclose in detail, describe the allowed scope and ask the employer how to provide any necessary evidence through an appropriate route. Do not invent a substitute story with different numbers and present it as an actual project. A hypothetical example should remain labeled hypothetical.
Keep your local application record proportionate. You may retain your original draft, final draft and factual notes without collecting unnecessary sensitive material. If you use a shared search tracker, store the application state and deadline there rather than confidential project details or personal assessment content.
Review the acknowledgment and final submitted text
Return to the actual application question after editing. Check the response length, requested focus and any required fields. A polished project paragraph may still fail to answer why you want the role or what experience the employer asked about. Relevance needs a separate review from grammar.
Read the AI-policy acknowledgment and any other certification independently. Confirm that you understand the policy you are acknowledging. Do not select a required “Yes” merely because the form cannot proceed without it if you have not read or cannot truthfully make the statement.
Inspect the final text in the application rather than assuming the editor’s version pasted correctly. Check for missing paragraphs, truncated sentences and unintended formatting. Open the selected resume and make sure it preserves the same supported facts. A corrected text answer should not coexist with an inflated claim in the attachment.
After the final action, retain the actual submission confirmation. This article’s review of a public form does not establish any candidate’s application state. Your receipt, employer communication and account record are the evidence relevant to your own application.
Questions to resolve before using AI
Can AI write the first draft for me? The checked Anthropic guidance asks you to create it yourself. Use your own experience and thinking as the starting point, then apply the permitted refinement rules.
Can a tool suggest a metric? Treat a proposed number as unsupported until your records establish it. Do not turn an editing suggestion into a factual claim.
Does application editing permission cover take-home work? No automatic extension follows. Read the task-specific instructions and obtain permission where required.
Must I submit every prompt I used? This guide does not establish such a blanket requirement. Follow actual disclosure questions and be truthful about your process.
What if I need an accommodation? Use Anthropic’s current careers-page process and the recruiting contact for the relevant step. Do not assume an unapproved assistance method is an accommodation.
When you compare fresh roles on LandOffer, save each employer’s assistance rules with the posting. Write your own first draft, verify the final claims and recheck the assistance rules before moving to an assessment or live interview.
Sources and Further Reading
- Anthropic: Guidance on Candidates’ AI Usage — Dated July 10, 2025; first-draft, refinement and stage-specific boundaries.
- Anthropic: Careers — Current policy entry point and applicant accommodation process.
- Anthropic: Staff+ Research Systems Engineering Application — One public role’s policy acknowledgment, observed October 12, 2026; not a universal application template.


