interview-sim

Practice the screen a company actually runs, on a real clock, in your own editor.

Paste a job posting. It researches what that company's technical screen is, writes an original question in that shape, proves the question can grade, and starts the clock. Real files, hidden tests, and a deadline that does not negotiate.

• Researches, never guesses. Live candidate reports, dated and sourced, every session. No stored company data to go stale
• Writes the question for the round. An original problem shaped by what the company reportedly asks, never a reused or real assessment item
• Proves it can grade before you sit it. Every question ships deliberately-wrong solutions; if the hidden tests miss one, the question is refused before the clock starts
• A clock that cannot be argued with. A script owns the time and the grades; a correct answer submitted one second late is refused
• Teaches after. report shows where the time went per question; debrief names what each failed hidden test was checking, with a reference solution on request

Install

Claude Code:

/plugin marketplace add chrisjacksonn/interview-sim
/plugin install interview-sim@interview-sim

Any agent that reads the Agent Skills convention:

npx skills add chrisjacksonn/interview-sim

Requirements: Python 3 (3.9+, stdlib only, nothing to pip install). For the fastest sitting (~2 minutes to the clock): /model sonnet, then /effort low.

Use

/interview-sim:sim https://jobs.example.com/some-posting

Or in plain words: "prep me for Ramp", "give me a GCA mock", "interview me". The agent researches, briefs you on what it found and how well sourced it is, writes the question, and the clock starts. Work in your own editor; submit tells you how many hidden tests passed and nothing else. Which ones failed is the exercise.

While you work: split your terminal and run ./timer for a live countdown (start writes that shim both into the workspace and beside interview-sim-sessions/, so it needs no setup and no cd). When it ends (or you end it): report breaks the clock down per question, and debrief turns the hidden tests into the lesson:

q1  Bracket Check   25/28   55:00 (79% of the sitting)
  - long unbalanced
  - one unclosed at the end
  - opener never closed

Three shapes, parameters yours to set:

How it stays honest

Every question must prove its tests can tell right from wrong. Questions ship with deliberately-wrong solutions, and each one must be caught by at least one hidden test. Shipped questions clear this in CI on every commit; questions written live clear the identical gate before the clock starts. A suite that passes everything grades nothing, and grading against one is worse than not grading at all.

There is no score. The scales real platforms report are proprietary and cannot be reproduced, so the report gives what was measured: questions solved, hidden tests passed, and where the time went. A question that timed out counts as zero rather than vanishing from the average.

Nothing is cached. A second session for the same company searches again, because hiring processes change between cycles and stored answers rot invisibly. Your own knowledge is the exception: notes in the practice directory, or a pasted invite email, are ground truth and skip the search.

Every question is original. Written to public format descriptions; found question text is only ever a signal about format and topics. PRs containing real assessment items are rejected.

The fine print, deliberately public: METHODOLOGY.md states exactly what is and is not measured, ETHICS.md covers what this tool refuses to be (the hidden tests are readable in this repo; they are hidden from the session, not from you), and skills/sim/references/formats.md documents the format specs questions are written to.

The clock, grading, and reports are plain Python; the agent is only needed for research and question-writing. sim is the shell function every session workspace README defines.

sim start --format gca          # a 4-question exam from the shipped set
sim status                      # time remaining
sim submit --question q1        # grade against the hidden suite
sim timer                       # live countdown, for a second pane
sim end                         # give up on purpose; opens the debrief
sim report                      # per-question results and time breakdown
sim debrief --question q1       # failed tests explained + reference solution
sim list / sim progress         # history, and what it adds up to
sim check                       # is this machine set up

Sessions are real directories under ./interview-sim-sessions/, one folder per company and round, each self-gitignored. ICA runs put four gated levels on one solution.py; every level re-runs the ones below it, so breaking level 1 while adding level 4 costs you.

A timed sitting stops being timed if every file read asks permission. Shift+tab toggles auto mode for a session, or scope it to your practice directory with .claude/settings.json:

{"permissions": {"defaultMode": "auto"}}

The skill mentions this once per session and cannot change permissions itself, deliberately.

Developed and tested in Claude Code on macOS; Linux works the same way. Windows falls back from process-group kills in the grader; check first. Codex, Cursor, and other Agent Skills hosts should run the same scripts, but no full session has been sat in one: treat as untested, and an issue saying whether it worked is genuinely useful.

Python is the graded language today. The engine opens no network connection; the agent driving it does, for research, exactly as if you searched yourself.

Contributing

Bug reports, platform reports, and feature PRs (issue first) are welcome. See CONTRIBUTING.md.

MIT. Not affiliated with any assessment platform or employer.