Here’s a thing that’s becoming more common: engineers showing up to system design interviews with an AI assistant running in the background. Not to cheat outright, but to have something that can surface a concept they’ve blanked on, confirm an estimation they’re unsure about, or prompt them when they’re stuck on trade-offs.
Whether that’s ethical is a separate conversation. This post is about what real-time AI assistance actually does well in a system design context, what it doesn’t do well, and how people are using it.
What “real-time AI help” even means here
There are a few different setups people use.
The most basic: a chat window open on a second monitor where you can type quick questions like “what are the trade-offs between Redis and Memcached for a session store?” and get an answer in a few seconds. Not ideal during a live interview because you have to look away and type.
More sophisticated tools listen to the audio of the call (with your permission, usually) and surface relevant prompts based on what’s being discussed. If the interviewer asks about rate limiting, the AI might show you a quick card with token bucket vs leaky bucket, when you’d use each, and common implementation pitfalls. You glance at it, it confirms what you already knew, and you keep talking.
That second mode is closer to what Craqly does: it follows the conversation and surfaces context-relevant prompts rather than waiting for you to ask. The difference matters because typing mid-interview is awkward and obvious, but glancing at a sidebar isn’t.
Where it actually helps
The honest answer is: in the gaps. Not the core of your answer, but the stuff you know but can’t access under pressure.
There’s a well-documented phenomenon where retrieval under stress is harder than retrieval in a calm study session. (The APA’s research on memory and stress covers this, though most engineers know it experientially from blanking on something obvious during a phone screen.) Real-time prompts work as retrieval cues. They remind you of things you know, rather than teaching you things you don’t.
That’s an important distinction. If you genuinely don’t understand consistent hashing, seeing a one-line prompt about it during an interview isn’t going to help you explain it convincingly. The interviewer will follow up. You’ll get stuck. AI assistance for system design works best when you’ve already done substantial prep and the prompts are filling gaps, not building foundations.
Specific things where it helps:
- Blanking on a term (what’s the word for when your cache invalidation strategy is “lazy” vs “eager”?)
- Estimation sanity checks (“is 35 writes per second reasonable for a 1% DAU posting rate at 300M users?”)
- Prompts when you’ve gone quiet for too long and need to get talking again
- Catching when you’ve forgotten to address non-functional requirements
Where it doesn’t help (and actually hurts)
Reading a suggested answer while talking is cognitively expensive. Your verbal reasoning and your reading are competing for the same working memory. I’ve seen demos where the person using the AI assistant ends up speaking in a slightly mechanical, stop-start rhythm as they parse the suggestion and integrate it. That’s worse than going blank for a second and recovering naturally.
The other failure mode: the AI suggests something technically correct but not appropriate for the constraint you’ve established. You said you’re designing for a small internal tool with 10,000 users, and the AI surfaces a Cassandra recommendation. You read it, say “we could use Cassandra here,” and the interviewer rightly asks why. You don’t have a good answer because you were following a prompt, not your own reasoning.
This is the version of AI assistance that actually undermines candidates. It creates a gap between what they say and what they understand. Interviewers find that gap quickly.
The prep question is more interesting than the interview question
Honestly, real-time AI assistance during prep is more valuable than during the actual interview. Running mock system design sessions where you get challenged on every decision, where the AI asks “you chose PostgreSQL here, but what happens at 500 million rows? what’s your sharding strategy?” that’s practicing the right skill. That’s building the muscle for narrating your reasoning.
The 2024 Stack Overflow Developer Survey found that 76% of professional developers are using or plan to use AI tools in their development workflow. The engineers who do well in system design interviews are, I suspect, the ones who’ve been using AI tools interactively during learning, not passively during testing.
There’s something worth saying here: a lot of the engineers I’ve talked to who tried AI assistance during live interviews came away thinking it helped less than they expected. Not because the AI gave bad answers, but because the cognitive overhead of monitoring it while managing a real-time conversation was higher than expected. The prep use case, where you can stop, read, question, and internalize, is where most people find the real value.
Practical setup if you want to try it
A few things that seem to make it less disruptive:
- Put the assistant on a second screen positioned slightly below your main screen rather than beside it. Down-and-back is less obviously not-looking-at-the-interviewer than left-right.
- Use an assistant that auto-surfaces prompts rather than requiring you to type queries. Typing mid-interview is the version that feels the most awkward and draws the most attention.
- Practice with the assistant during mock sessions before using it in a real interview. If you’ve never used it while talking through a design, the first time in a real interview is the wrong time to learn the workflow.
- Set a personal rule: don’t read suggestions that are more than one sentence. Long suggestions require reading time you don’t have.
Craqly’s real-time interview mode is built around this use case specifically: short, context-triggered prompts rather than full answers. The assumption is that you know the material and need cues, not explanations.
What this means for how you prep
If you’re going to use any AI assistance during a system design interview, the single best thing you can do beforehand is take at least a dozen mock sessions where you explain your decisions out loud, get challenged, and practice staying coherent under pushback. That skill is what the interviewer is actually scoring. The AI can help you remember the term “idempotency key” when you blank on it. It can’t help you explain why you need one.
Do the prep first. The assistance is a safety net, not a substitute. And safety nets don’t help if you’ve never learned to walk the wire.