Interview Coach
Turns Claude into a full-lifecycle job-search coach: from JD decoding and resume optimization through mock interviews to offer negotiation, scoring every answer and tracking your weak spots over time.
Pure prompt-based skill: no shell commands, network calls, or dangerous permissions; writes only a local coaching_state.md with transparent data flow. Deductions: persists sensitive personal data (resume, job-search status, interview feedback) with no stated retention scope, no deletion/rollback mechanism, silent mid-session state writes, incomplete confirmation flow; publisher unverified, attribution limited.
Internal rules are self-consistent with clear priorities and confidence labels instead of fabricated certainty. Deductions: behavior depends on ~30 reference files not present in the reviewed material, so key paths cannot be statically confirmed; no tests, no error-handling spec, undefined behavior on corrupted state files; static cap prevents scoring above 10.
Audience, scenarios, command registry, mode-detection triggers, and declared non-fit boundaries (Rule 11: communication coaching, not domain evaluation) are clearly written. Deductions: no Chinese-language support evidence; examples are US-centric; requires paid Claude/Codex plans, with reachability and cost for mainland-China users undiscussed.
Well-layered docs with progressive disclosure, a coherent VERSIONS.md roadmap with shipped versions, and a complete MIT license. Deductions: maintenance responsibility and update path rest on a single-author repo with no per-version changelog or centralized known-limitations disclosure; the referenced files were not available for static confirmation.
README workflow examples and expected outputs are concrete and the design (state continuity, calibration loop) is substantive. Deductions: static review cannot verify any output's actual usability; core claims (scoring accuracy, transcript format detection) lack execution evidence, so the static cap of 7 is not reached given thin evidence.
SKILL.md, README, and VERSIONS.md are mutually consistent auditable primary material. Deductions: all effectiveness and 'gets sharper with use' claims are author self-report; no test suite, no CI, no third-party execution evidence, no independent reproduction, and facts/inferences are not strictly separated; static cap is 5 and evidence is thinner than that.
- This is a static review only; no commands were executed. The ~30 referenced files were not part of the reviewed material, so real behavior may differ from descriptions.
- coaching_state.md persists sensitive personal data (resume, interview logs, rejection feedback); manage storage and deletion of this file yourself.
- The skill requires a paid Claude/ChatGPT plan and shows no Chinese-language support; mainland-China users face reachability and language risks.
- Scoring calibration and company-research outputs are independently unverified; verify company-specific claims (culture, process, values) yourself.
What does this skill do, and when should you use it?
Interview Coach is a Claude Code-based interview coaching skill covering the full job search lifecycle with 23 commands across application materials, interview prep, practice, and post-interview analysis. It scores every answer on five dimensions (Substance, Structure, Relevance, Credibility, Differentiation), calibrated to your seniority, then uses a decision tree to triage your bottleneck toward targeted drills. A persistent coaching_state.md file maintains memory across sessions, so the system gets sharper the more you use it. It supports PM, Engineering, Design, Data Science, Research, Marketing, and Operations roles, and handles behavioral, system design, case study, panel, and mixed interview formats.
On kickoff it reads your resume and target role to build a coaching profile. For transcripts, it auto-detects the source format (Otter, Zoom, Grain, Teams, etc.) and scores per question or per design phase. It builds and manages a storybank (STAR text, earned secrets, rapid-retrieval drills, portfolio-optimized story-to-question mapping). The research and prep commands run company research with every claim tagged to a source tier (verified, general knowledge, unknown). It runs an 8-stage drill progression plus full 4-6 question mock interviews, and progress tracks trends, self-assessment deltas, and scoring drift against real interview outcomes. Separate commands handle LinkedIn, resume, pitch, outreach, JD decoding, presentation rounds, early-process comp, and post-offer negotiation. All state auto-saves to coaching_state.md so sessions can be interrupted and resumed.
- A candidate running a multi-week job search who wants systematic storybank management, transcript reviews, and tracking across company interview loops
- A job seeker with an interview days away who picks the Quick Prep track for fast company research, a prep brief, and focused drills
- Someone fresh out of a real interview who runs debrief the same day to capture questions and interviewer signals, then analyze on the transcript
- A candidate with a formal offer who uses negotiate for strategy and exact scripts
- A candidate facing the recruiter-screen salary question who uses salary to build a defensible range and stage-specific scripts beforehand
- An applicant deciding where to invest effort who uses decode to analyze a JD or batch-compare 2-5 job descriptions
What are this skill's strengths and limitations?
- 23 commands covering the full lifecycle from JD decoding to offer negotiation — not a generic question bank
- Persistent state file delivers genuine cross-session continuity; scores, storybank, and interview intelligence accumulate over time
- Five-dimension scoring plus root-cause diagnosis and decision-tree triage means different candidates get different paths
- Strict evidence standard: company-specific claims are tiered as verified, general knowledge, or unknown, reducing hallucinated advice
- Auto-detects multiple transcript formats with dedicated analysis logic for behavioral, system design, panel, and mixed formats
- MIT licensed, with both Claude Code and OpenAI Codex usage paths documented
- Requires a paid Claude or ChatGPT subscription and an environment that can read/write local files
- Output quality depends heavily on input quality — real resumes, full JDs, and real transcripts; thin inputs yield limited coaching
- Company research depends on the model's web retrieval and source reliability; the repo ships no automated test suite
- For system design and case study formats, the coaching value is communication-level rather than domain expertise — a boundary the skill itself states
- The 23-command system plus reference files has a real learning curve for new users
How do you install this skill?
1) Clone the repo: git clone https://github.com/noamseg/interview-coach-skill.git and cd into it. 2) For Claude Code, run mv SKILL.md CLAUDE.md (same for other file-system-access environments like Cursor); for OpenAI Codex, run mv SKILL.md AGENTS.md. 3) Open the folder in Claude Code or Codex. Requires a paid Claude or ChatGPT plan. No other installation methods are documented.
How do you use this skill?
Type kickoff in the conversation and share your resume, target role, and timeline; the system builds your profile and gives a prioritized action plan. Then use commands as needed: research Notion for company research; prep Stripe for a role-specific prep brief; analyze followed by a pasted transcript for per-question scoring and diagnosis; practice or mock behavioral Stripe for drills and full simulations; stories to manage the storybank; progress for trends and calibration. Every workflow ends with a state-aware next-step recommendation. For maximum candor, set feedback directness to 5 during kickoff to activate the Challenge Protocol.
How does this skill compare with similar options?
The README explicitly contrasts this skill with asking ChatGPT for interview help: generic LLM advice is identical for everyone, while this system scores over time, diagnoses root causes, and adapts based on data — the difference between a textbook and a coach.