Reviewing a single developer take-home assignment takes between 15 and 30 minutes on average. Across a pipeline of 20 candidates, that's 5–10 hours of senior-engineer time per role — time that almost never converts into a better hiring signal than the first 5 minutes did.
This guide walks through what a take-home review should actually check, in what order, and where automation belongs.
Why take-home reviews eat your week
Most reviewers fall into the same trap: they open the repo, start reading code top-to-bottom, and form an opinion in the first 90 seconds. Everything after that becomes confirmation. The actual signal — did the candidate meet the requirements? — gets answered by skimming README.md, not by reading code.
The four checks that matter:
- Requirements coverage — did they do every item from the brief?
- Code quality vs. requirement fulfillment — was it shipped or shipped-well?
- AI-written code detection — did they actually write this?
- Whether the app actually runs — does
npm install && npm startboot a working server?
Everything else (commit hygiene, file organisation, naming) is a tie-breaker.
A rubric you can copy-paste
Here's the rubric we use at CodeVerdict. Five dimensions, weighted, with explicit hire/no-hire thresholds.
| Dimension | Weight | What you check |
|---|---|---|
| Requirements met | 35% | Each line item in the brief mapped to a commit or file |
| Code quality | 20% | Readability, error handling, abstraction level |
| Tests | 15% | Coverage of happy + edge paths, not just describe('it works') |
| Security | 15% | Hardcoded secrets, SQL injection patterns, missing input validation |
| AI-written code | 15% | Token perplexity, naming entropy, commit history |
Hiring thresholds:
- Strong hire: ≥ 85, no security issues, AI score < 30
- Hire: 70–84, at most one security issue
- No hire: < 70 OR any security score < 40 OR AI score > 80
The case for automation
A senior engineer making $200K costs your company roughly $100/hour fully loaded. At 20 minutes per review, you're spending $33 of senior engineering time on each candidate — most of whom you'll reject. Across a year of hiring, that's tens of thousands of dollars of opportunity cost burned on a task that's mostly mechanical.
Automation isn't about replacing the engineer's judgment. It's about handing them a pre-scored report so they spend 5 minutes on the interesting questions instead of 25 minutes on requirement-checking.
How to use automated review with human judgment
The 90/10 split:
- 90% automated: requirements mapping, code execution, security scan, AI detection, basic quality metrics.
- 10% human: subjective architectural choices, communication quality, the "would I want to work with this person" check.
Run the automated review first. If it's a No hire for objective reasons (failed requirements, security issues, app doesn't run), stop. If it's a Hire or Strong hire, spend your 5 minutes on the architectural read and the interview question selection.
Frequently asked questions
How long should a take-home assignment take a candidate? Aim for 2–4 hours of focused work. Longer than 4 hours and your pass rate craters because candidates who already have offers self-select out. Be explicit about the time budget in the brief.
Should I tell candidates I'm using an AI tool to review? Yes. Transparency increases trust and discourages candidates from using AI themselves (since they know it'll be detected). A line in the brief like "Submissions are scored automatically for requirement coverage, AI-written code, and execution" is enough.
Can AI detect AI-written code reliably? Yes, with caveats. The signal is statistical — token-level perplexity, naming entropy, and commit history patterns. A score of 80+ is a strong indicator. A score of 60–80 is grey-zone (probably AI-assisted, possibly heavily styled human code). Below 60 is almost certainly human.
Are take-home assignments still worth it in 2026? Yes, if you grade them in under 5 minutes per candidate. The signal-to-time ratio is the highest of any hiring stage when you remove the manual review overhead.
Want to skip the manual rubric entirely? Try CodeVerdict on your next assignment — it generates this exact scored report in 60 seconds. No signup needed.