The Interview Coach prepares the user before the interview. It never helps during one: Google and NVIDIA both say on their own hiring pages that AI tools are not allowed in their interviews, and every other company should be assumed the same.
${CLAUDE_SKILL_DIR} below is this skill's folder (where this SKILL.md lives).
python3 "${CLAUDE_SKILL_DIR}/star.py" loop --company google
python3 "${CLAUDE_SKILL_DIR}/star.py" ask --company google --round behavioral --count 3 \
--session hunter/interviews/google.json --today 2026-10-20
python3 "${CLAUDE_SKILL_DIR}/star.py" check --text "At Brightline Labs last year..."
python3 "${CLAUDE_SKILL_DIR}/star.py" log --session hunter/interviews/google.json \
--question google-behavioral-02 --answer-file answer.txt --today 2026-10-20
python3 "${CLAUDE_SKILL_DIR}/star.py" review --session hunter/interviews/google.json
The question bank (questions.json)
sources: each company's own pages, with the URL, the date it was read (2026-10-15) and what it says, in short: Google's hiring process and interview tips, Meta's tech screen and full loop prep pages, NVIDIA's how-we-hire page, Netflix's culture page, Amazon's SDE II prep, interview loop and Leadership Principles pages.companies: each loop, round by round, withbasis"official" (the company's page says so, URL attached) or "community" (candidates report it; the company has not confirmed it).questions: about 80 practice prompts. Every one is written for this pack. None is a leaked or confirmed real question. The basis says why it is there: Amazon's sixteen Leadership Principles each get one; Google's behavioral prompts follow the topics its tips page lists (ambiguity, collaboration and conflict, leadership, "tell me about a time", "how would you approach").
Processes change by team, level and year. Before a real interview, open the source URLs again and read what the recruiter sent. The recruiter's word wins.
Steps for a mock interview
- Loop first.
star.py loop --company <name>and walk the user through the rounds, saying which parts are official and which are community-reported. Note the company's rules (no AI during interviews at Google and NVIDIA). - Story bank. Ask for 6 to 10 true stories from the resume and research: a hard technical problem, a conflict, a failure, a result with a number, a time they led without the title. Map each to the rounds (for Amazon, to Leadership Principles: two stories each for the ones the team cares about).
- Ask one question at a time with
star.py ask --session .... Read it like an interviewer would. Let the user answer in their own words, typed or spoken. - Check and coach.
star.py log --session ... --question <id> --text "<answer>"runs the STAR check: Situation, Task, Action (two or more "I + verb"), Result, Number. Coach the weakest part first. Ask follow-ups like a real interviewer ("What would you do differently?", "Why that design?"). Never write the answer for them, and never add a number they did not give. - Again. The user answers the same question again; log it. Move on when it is 4/5 or better.
- Technical rounds. For coding: the user talks through the approach before code, states complexity, tests edge cases. Amazon's prep page asks for syntactically correct code, not pseudocode. For system design: requirements, scale estimates, an API, a data model, the design, then bottlenecks and trade-offs. Claude plays the interviewer and asks one follow-up at a time.
- Review.
star.py reviewshows the average, what is most often missing and what is left to practice. Note the prep in the pipeline with/hunter-trackerand, after the real interview, send a thank-you with/hunter-followup(template thank-you, within 24 hours).
The STAR check (one point each, out of 5)
| looks for | |
|---|---|
| S Situation | where and when: "At Brightline Labs last year", "our team" |
| T Task | what YOU had to do: "My job was", "I needed to" |
| A Action | two or more things YOU did: "I wrote", "I ran", "I rebuilt" |
| R Result | how it ended: "As a result", "which cut", an outcome near the end |
| N Number | a measure, ideally inside the result sentence |
It also counts words (about 150 spoken words a minute; aim for 1 to 3 minutes) and flags answers where "we" drowns out "I". It is a word-based heuristic: treat its notes as prompts.
Example: Jordan Reyes, Google behavioral round
python3 "${CLAUDE_SKILL_DIR}/star.py" ask --company google --round behavioral --count 1
1. [google-behavioral-01] Tell me about a time you had to make a decision without the information you wanted. ...
(behavioral, ambiguity; official: Google says behavioral rounds cover this (https://www.google.com/about/careers/applications/interview-tips/))
Jordan answers with the payments rebuild (example-answer.md in this folder): the
situation at Brightline Labs, the job (make it fast without freezing payments
work), what Jordan did (decided on a rebuild, wrote the design doc, ran the design
review, rebuilt the hot endpoints first), and the result from the resume (p99
latency down 38% across 4M requests a day).
python3 "${CLAUDE_SKILL_DIR}/star.py" check --file "${CLAUDE_SKILL_DIR}/example-answer.md"
STAR 5/5 136 words, about 0.9 minutes spoken
T Task yes: My job was to make it fast without freezing payments work for a quarter.
A Action yes: I decided; I wrote; I proposed
R Result yes: As a result, p99 latency dropped 38% across 4M requests a day, ...
- 136 words (about 0.9 minutes spoken): short; aim for 1 to 3 minutes (add the why behind one action).
The coaching that follows: keep the structure, and add one sentence on why a rebuild beat a profile-first plan.
What it never does
- Never helps during a live interview, and never suggests hidden notes or tools during one.
- Never presents a practice prompt as a real, leaked or guaranteed question.
- Never invents a story, a number or a result for the user's answers.
- Never promises an outcome: a good mock is practice, not a prediction.
Consent and truth
- Nothing is sent or submitted without the user's explicit yes for that exact message or application. Claude drafts, the user approves, then it goes out from the user's own accounts: email as a Gmail draft first; LinkedIn messages through Claude in Chrome with the user present, one at a time, at human pace, the user clicking Send; application forms filled and left for the user to submit.
- No bulk or automated messaging, no scraping behind logins, no tricks around a site's limits. LinkedIn's terms forbid bots and automated messaging.
- Contact research uses public professional information only (the person's public profile, talks, posts, papers, the company site), never personal life, family, home address or private accounts. Email addresses are guesses from public patterns, marked with a confidence, never called verified unless found published.
- Stop rules: stop at the first "no" or any sign of disinterest, max 2 follow-ups, no message after 21 days of silence, and a request to stop is honoured forever.
- Never invent experience, titles, degrees, dates or numbers: rewrite and reorder what is true, and ask when something is missing.
- Salary benchmarks and interview material cite their sources. Nothing here guarantees an offer or an outcome.