TL;DR
- Whiteboard interviews over-index on memorization and under-measure real engineering judgment.
- Modern engineering is AI-assisted; interviews should be, too.
- The best signal now is how someone plans, prompts, verifies, and debugs with AI inside a real workflow.
- AI-assisted work-sample interviews produce clearer evidence than "solve this on a whiteboard."
Y Combinator (YC) — the legendary startup accelerator behind companies like Airbnb, Stripe, and Dropbox — sparked debate across tech this month with a single update to its application process.
In a LinkedIn post, YC announced applicants will now be asked to upload a transcript from an AI coding tool. YC wants to see how founders "plan, design, debug, and ship" their most important features, according to the post.
The response was emphatic and largely supportive.
"This is such a smart move," wrote Prateek J. in the comments. "The way someone uses AI tools reveals so much more about their engineering judgment than just the final product."
"Finally capturing the how, not just the what. Well done," wrote Subhav G.
The update from one of tech’s most influential institutions reflects a position Howdy has held for years. Today, how engineers think with AI matters more than how they perform without it.
Our interviews have long included live collaboration with tools like ChatGPT and Copilot. We review how engineers structure prompts, refine outputs, and debug generated code in real time.
Why whiteboard interviews don’t predict job performance anymore
For decades, technical interviews leaned on whiteboard exercises: solve algorithmic problems from memory, under time pressure, often without documentation, tools, or a real development environment.
That format had benefits. It was standardized, fast to administer, and easy to compare across candidates.
But it’s increasingly misaligned with the job.
Modern engineering work happens inside:
- Complex, evolving codebases
- Distributed teams and async collaboration
- Real constraints (tests, CI, performance, security, maintainability)
- AI copilots used for research, scaffolding, debugging, refactoring, and documentation
The differentiator is no longer recall. It’s judgment inside an AI-assisted workflow: framing the problem, managing context, verifying outputs, and making tradeoffs responsibly.
What to replace whiteboard interviews with: AI-assisted work-sample interviews
An AI-assisted work-sample interview evaluates a candidate inside a realistic environment, with the same tools used on the job (including AI), and produces evidence of how the candidate thinks.
Instead of “can this person remember the right algorithm,” the question becomes:
- Can this person define the objective and constraints?
- Can this person use AI effectively without outsourcing responsibility?
- Can this person verify correctness and catch subtle failures?
- Can this person communicate tradeoffs and reasoning clearly?
The output isn’t just code. It’s a trail of decisions: prompts, iterations, tests, diffs, and explanations.
How Howdy interviews engineers with AI tools
At Howdy, AI is part of the interview environment because it’s part of the job.
Candidates work inside realistic repositories with access to tools like ChatGPT or Copilot. The interview is designed to surface how someone:
- Frames the business objective before writing code
- Structures context and constraints
- Prompts and iterates intentionally (not randomly)
- Reviews and verifies AI-generated output
- Debugs when the model introduces subtle flaws
- Communicates decisions and tradeoffs
A transcript (or prompt history) becomes evidence of reasoning in motion.
Strong, AI-enabled engineers treat AI as leverage that requires oversight. They move quickly, but deliberately. They understand that output still carries human responsibility.
What “good” looks like (and what it doesn’t)
Senior engineers reveal themselves before implementation begins. The best signal often shows up in the first 10 minutes: framing, constraints, and approach.
1) Clarity before code
Strong candidates articulate:
- The goal in plain language
- Constraints (time, performance, security, maintainability)
- Edge cases and failure modes
- A plan of attack before generating solutions
Clear framing leads to cleaner prompts. Cleaner prompts lead to stronger outcomes.
2) Context is the multiplier
AI tools amplify whatever context they’re given.
Engineers who understand this treat context as a first-class input:
- They explain architecture before modifying it
- They identify dependencies before introducing change
- They communicate assumptions explicitly
- They keep the model grounded in the repo’s reality (not generic patterns)
Software is interconnected. AI magnifies both clarity and confusion.
3) Verification is the job
The best engineers don’t "trust the model." They verify.
They:
- Run tests (or write them)
- Check edge cases
- Validate performance implications
- Review diffs like a code reviewer, not like a spectator
- Catch subtle issues (off-by-one logic, incorrect assumptions, security footguns)
4) Process over performance
Code generation is one stage in a larger workflow.
A strong interview surfaces end-to-end thinking:
- Reasoning under constraint
- Progressive refinement
- Consistency across stages
- System design tradeoffs and decision defense
As Frank Licea, CTO at Howdy, explains:
"Strong engineers code review is the bottleneck rather than code review. The best ones focus on everything that happens before and after the code is written."
Howdy hires for the thinking that surrounds the code.
| Competency | What to look for | Red flags | Evidence | |
| Problem framing | Restates objective, clarifies requirements, defines “done” | Jumps into code immediately | Notes, initial plan, questions asked | |
| Context management | Provides repo context, constraints, dependencies | Prompts are vague; ignores existing patterns | Prompt history, architecture explanation | |
| Prompt quality | Specific, iterative, grounded in codebase | “Do it all” prompts; random thrashing | Transcript, iteration trail | |
| Verification | Tests, edge cases, careful review | Accepts output without checking | Test runs, diffs, review comments | |
| Debugging | Systematic isolation, hypothesis-driven | Trial-and-error without learning | Debug steps, logs, reasoning | |
| Communication | Explains tradeoffs, narrates decisions | Can’t justify choices | Debrief, written summary |
How to run an AI-assisted interview (step-by-step)
- Choose a realistic task (bug fix, feature slice, refactor, integration)
- Define constraints and success criteria (tests passing, performance, security, style)
- Allow AI tools (ChatGPT, Copilot, Claude Code, etc.) and require transparency
- Observe context setup (how the candidate explains the repo/problem to the model)
- Require verification (tests, edge cases, review of generated code)
- Review the diff and reasoning (not just the final output)
- Debrief tradeoffs (what was chosen, what was deferred, what risks remain)
From screening to system: What interviews create downstream
The way talent gets evaluated shapes the kind of team that gets built.
At Howdy, the philosophy behind interviews carries into how engineers are trained and supported after placement:
- Sourcing top engineering talent across Latin America
- Rigorous vetting for technical depth and intellectual flexibility
- Structured AI training embedded into real workflows
- Ongoing mentorship, architecture reviews, and hands-on experimentation
Fewer than 5% of applicants make it through. Developers who completed Howdy’s AI training pilot saw a 52% increase in productivity, and many go on to lead AI adoption within client organizations.
FAQ: AI in technical interviews
Should candidates be allowed to use ChatGPT or Copilot in interviews?
Yes, if the job allows it. The interview should measure how someone uses AI responsibly, not whether they can pretend it doesn’t exist.
How is "cheating" prevented if AI tools are allowed?
The goal isn’t to block AI. The goal is to evaluate judgment. Requiring a transcript, reviewing reasoning, and using repo-specific tasks makes it hard to fake competence.
What gets evaluated if AI writes some of the code?
Planning, context management, verification, debugging, and tradeoffs. Those are the skills that determine whether AI output becomes production-quality work.
What does a strong prompt workflow look like?
Clear objective, relevant repo context, constraints, incremental steps, and explicit verification. Strong candidates treat prompts like engineering artifacts.
Are whiteboard interviews ever useful?
They can be useful for specific roles or early screening, but they’re a weak proxy for real-world performance in modern, AI-assisted engineering environments.
What’s the best alternative to LeetCode-style interviews?
A realistic work sample inside a repo, with AI tools allowed, scored with a rubric that emphasizes verification and decision-making.
Howdy: Built for this era
The public conversation is catching up to a reality that has already reshaped high-performing teams.
AI accelerates production. Human judgment safeguards quality. The whiteboard interview belongs to an earlier chapter of engineering.
Howdy has been building for this one.
Ready to hire engineers who can think clearly, collaborate with AI responsibly, and ship at modern speed? Book a demo.