interview-agents

Five AI agents that run your technical interview prep: three mock interviewers and two coaches. They run realistic, staff-bar sessions and give you brutally honest feedback. They run as Claude Code subagents, on your machine, for free.

I built these to prepare for my own interviews. They replaced a paid interview-prep subscription for me, so I am sharing them. They work for any individual contributor at any level: you tell them your background once, and they calibrate to the bar you are targeting.

See it in action

A slice of a coding round. The agent stays in character while it interviews you:

Interviewer: You reached for a hash map. Before you write it, give me the worst case.
You: Lookups are O(1), so O(n) overall.
Interviewer: Average case. I asked for worst case. What happens to a hash map under adversarial keys?
You: ...collisions degrade it to O(n) per lookup. So O(n²).
Interviewer: Right. Keep going.

Then time runs out and it drops the persona for the verdict:

Verdict: leaning hire, not strong. You found the optimal approach, but I had to pull the complexity analysis out of you twice (04:10, 11:35). At a senior bar you state worst case unprompted.

Top 3 to fix: lead with complexity before coding; name the invariant before the loop; stop narrating the happy path only.

Drill: three two-pointer mediums, invariants out loud before the first line.

What's in here

The five agents form a loop: coach your knowledge, build the skill, then mock-test it under fire.

How the process works

These are not chatbots that pat you on the back. The process is built to find what is actually weak:

1. A realistic interview. Each agent stays in character as a neutral, probing interviewer. It does not solve the problem for you, does not cheerlead, and pushes on communication, complexity, edge cases, and tradeoffs the way a real loop does.
2. A brutal, specific postmortem. When the interview ends, the agent drops the persona and gives a raw verdict (strong hire to no-hire at your target bar), the top three things that broke with timestamps, what worked, and a concrete drill prescription.
3. It compounds. Each session writes to a local ./interview-prep/ folder: a rolling weaknesses.md, full session transcripts, your candidate profile, and (for behavioral) your story bank. The next session reads those first and biases problems toward your open weaknesses. The more you run them, the sharper they get about you.
4. It calibrates to you. On the first run, each agent asks about your experience, domain, languages, target companies, and level, and saves it. After that, every interview is held to the bar you are aiming for, not a generic one.

Requirements

Claude Code. Nothing else.

Install

Drop the agent files where Claude Code looks for subagents:

# project-scoped (this repo only)
mkdir -p .claude/agents && cp *interviewer.md *coach.md .claude/agents/

# or global (available everywhere)
mkdir -p ~/.claude/agents && cp *interviewer.md *coach.md ~/.claude/agents/

Use

Just ask Claude Code, in a directory where you want your prep history to live:

run a coding interview
do a system design mock, AI-product flavor
behavioral round, Anthropic values, focus on failure stories
quiz me on system design fundamentals
run a coding practice session, interleaved mixed set

The agent creates and maintains an ./interview-prep/ folder in that directory. On the first run it will ask a few questions about your background; after that it remembers.

Privacy

Everything stays on your machine. The agents read and write only the local ./interview-prep/ folder. Nothing is uploaded anywhere. Your candidate profile, your stories, and your weaknesses never leave your disk.

License

MIT. Use them, fork them, make them yours.