interview-gen

Generate spec-driven, take-home interview questions for software engineers. Each question is a runnable repository with open decisions, a decision log, a rubric, and trace-evaluable signal.

What it does

The skill generates interview questions in three steps:

1. Design — You describe your hiring need (role, format, domain, competencies), and the skill designs a scenario with a milestone ladder and open decisions
2. Build — The skill produces a runnable starter repo and candidate-facing materials (SPEC.md, AGENTS.md, CLAUDE.md, README.md)
3. Package — The skill writes a rubric.yaml for evaluation, a meta.yaml with question metadata, and a question.json file containing the full question in shareable format

Installation

Copy the interview-gen folder into your Claude skills directory:

# For Claude Desktop / Claude Code
cp -r interview-gen ~/.claude/skills/

# For other Claude setups
cp -r interview-gen <your-claude-skills-path>/

Usage

In Claude, invoke the skill by typing /interview-gen or "generate an interview question".

You'll be prompted for:

• role — hiring role (choose from available profiles: fe-react, generic-swe)
• format — screening, live-coding, or take-home
• mode — scratch (build from spec) or oss-derived (wrap an existing OSS repo)
• competencies — what you're evaluating (e.g., ["architecture", "refactoring", "integration"])

The skill then outputs a folder (<question-id>/) containing:

• meta.yaml — question metadata and prose (scenario, tagline, candidate_brief)
• candidate/repo/ — the git repo candidates receive
• review/rubric.yaml — evaluation rubric
• question.json — self-contained question format (see below)

question.json — shareable format

After generating a question, the question.json file contains the full question in a portable, application-agnostic format. You can:

• Share with your team — paste into CredWork to create a new inteview question
• Iterate locally — re-run the skill to regenerate question.json if you tweak the rubric or spec
• Version control — commit question.json to track the question definition over time

The schema includes all candidate-facing content (prompt, repo link if published), rubric criteria, competencies, and metadata.

Profiles

The skill ships with two role profiles:

fe-react

Front-end React engineer. Covers:

• React 18+, TypeScript, Vite
• State management, data fetching, caching
• Component design and accessibility
• Bundled local API server for realistic seams

generic-swe

General software engineer. A role-agnostic template suitable for:

• Backend, systems, or full-stack questions
• Questions where the role doesn't fit a specialized profile
• Custom stacks (document your own assumptions in the spec)

Core principles

• Spec-driven — Candidates receive a SPEC.md with open decisions, not a menu. They decide; the coding agent raises questions.
• Decision-focused — The rubric scores judgment (how they resolved conflicts, what they chose to build/skip), not just correctness.
• Trace-evaluable — A Decision Log captures every major choice and reasoning in real-time. Reviewers score from the trace + diff, not from hand-written summaries.
• No tests required — Candidates don't write tests; the rubric scores behavior by running the tool/app. Avoids circular validation.
• Runnable end-to-end — Fresh laptop + stated prereqs + README steps = running system. No cloud accounts, no magic.

For reviewers

The skill does not include review tooling. To evaluate submissions, you'll need to:

1. Receive the candidate's decision log, code, and any trace material
2. Score against the rubric.yaml using your organization's review process
3. Record outcomes: strong | adequate | weak | insufficient_evidence for each rubric item

See SKILL.md for rubric design principles and scoring guidance.

License

[Add your license here — e.g., MIT, Apache 2.0]

Questions?

See SKILL.md for detailed specification design rules, rubric contracts, pipeline steps, and examples.