The hiring challenge for AI-native operators

Single Grain is looking for people who can use AI as leverage and still beat a generic AI answer with judgment, proof, and execution.

Beat Claude is the public hiring challenge that runs through /apply and our careers page. It is separate from any internal AI fluency system at Single Grain.

1. Pick a challenge from /challenges
2. Read the brief
3. Build the required operating artifact
4. Submit your answer and evidence packet
5. We compare it to Claude's baseline in blind review
6. Strong submissions may get a private curveball or live walkthrough
7. If you beat Claude with evidence, we want to talk

AI can produce competent knowledge work quickly. We are looking for people who can do what AI still struggles with: judgment under constraints, creative leaps, practical execution, source checking, and knowing what should stay human.

If you only paste the brief into an AI tool, you will usually get something close to the baseline we already have. That means you tie. Ties do not advance.

Use whatever tools you want: AI, research, friends, spreadsheets, code, workflows, prototypes, or live data.

To beat Claude, you need to add something it could not reliably produce from the brief alone:

• A real operating artifact.
• Source-labeled numbers.
• Evidence that you tested, built, validated, or modeled something.
• Clear tradeoffs and failure handling.
• A sharp explanation of what AI should not do.

Time commitment: Each challenge is designed to take 1-2 hours. Written submissions are capped at 4 pages unless the brief says otherwise. Artifact links, code, sheets, dashboards, and short demos can be included separately.

The public scoring guide is intentionally high level. We do not publish detailed scoring keys, exact point bands, or all reviewer prompts because that rewards rubric gaming.

Review may include:

• Blind comparison against Claude's baseline.
• A private curveball after submission.
• A short live walkthrough of your artifact.
• A proof check for numbers, source records, logs, or assumptions.
• A second reviewer when decisions are close.

See SCORING.md for the public guide.

Every challenge submission should include:

1. Written answer: the strategy, plan, design, or build explanation.
2. Operating artifact: a sheet, repo, workflow, Loom, dashboard, scorecard, process doc, prototype, log export, or other artifact a reviewer can inspect.
3. Evidence log: proof tier for each major claim, using the tiers in SCORING.md.
4. Number source labels: every number labeled as observed, estimated, benchmarked, or assumed.
5. AI usage disclosure: tools used, what they did, what you changed, and what you checked.
6. Failure handling: what breaks the plan or artifact, how you would detect it, and what stays human.

Submissions without an operating artifact or source-labeled numbers are unlikely to advance.

Don't see your role below? Start with the General Challenge or the Internship Challenge.

Apply through our careers page: singlegrain.com/careers

Upload your challenge answer and include links to any artifacts.

1. Your submission is anonymized
2. A reviewer scores your answer alongside Claude's baseline
3. We check evidence quality, source labels, and artifacts
4. Strong submissions may receive a curveball or live walkthrough
5. If you beat Claude, we reach out to schedule a conversation

Top performers who beat Claude: Hall of Fame

Single Grain is a digital marketing agency working with companies like Amazon, Uber, and Salesforce. We're building an AI-forward team that uses technology as leverage, not as a crutch.

Open an issue or apply at singlegrain.com/careers

Inspired by Anthropic's performance take-home