talalahmed311/cornerman

Cornerman 🥊

A sparring partner, not a cheerleader.

A Claude Skill that runs a realistic, adaptive technical mock interview against a specific job. Multi-phase, turn-by-turn, with an honest scorecard and study plan at the end.

---

What it does

Give it your resume and a job description. It will:

1. Read the resume, classify your seniority (junior / mid / senior / staff), and build an interview plan.
2. Run a project deep-dive — probes for real impact first, drops into technical fundamentals when the impact story is thin.
3. Bridge your background to the target JD’s actual requirements.
4. Give you a coding problem (async take-home) from the Blind 75, sized to your level. You solve it, paste your code, and Cornerman walks through your choices — correctness, complexity, brute-force vs. optimal.
5. Switch into a staff-architect persona for a system design (HLD, async take-home) — you sketch the design (whiteboard photo, Excalidraw, screenshot, whatever), upload it with a short written approach, and then walk through it in a live Q&A.
6. Zoom into the same system for low-level design (LLD, live) — data model, APIs, concurrency, edge cases.
7. Deliver a structured scorecard, honest hire signal, and study plan.

Expect deliberate pressure and evidence-tied feedback. That’s the point — that’s what a cornerman does between rounds.

Async take-home phases

Two phases (coding and system design) run as take-home rather than live:

• Setup — Cornerman poses the problem and answers your clarifying questions with concrete numbers (e.g. “10M DAU, 100:1 read/write, p99 under 200ms”). Not “you decide” — real interviewer numbers.
• Solve — Cornerman records the start time to the MCP, schedules an auto-grade callback at your time budget (via the scheduled-tasks MCP), and waits silently. No hints.
• Submit — you paste code (coding) or upload a diagram + written explanation (system design). Cornerman cancels the auto-grade callback and grades the real submission.
• Timeout branch — if you never return, the scheduled callback fires in a fresh conversation, loads your session from disk, records the no-submit, and saves a floor score. You lose that round; the rest of the interview still stands.
• Walkthrough — after the grade, Cornerman runs a live Q&A on your specific choices.

How to invoke

In a Claude conversation, provide:

• Your resume — upload a PDF/DOCX or paste the text.
• The target role — a JD text, a job posting URL, or a company URL + role title.
• Say something like “cornerman, interview me for this role”, “run a mock interview”, or “prep me for this JD”.

Cornerman will trigger and start with intake.

Optional at intake

• Session mode — full run (default) or single-phase (“just do system design”, “just LLD on my payments project”).
• Harshness dial — supportive coach, standard interviewer (default), or brutal bar-raiser. Changes tone, not scores.

What to expect

• One question per turn. Cornerman asks, then waits. Answer as you would in a real interview.
• Phase handoffs are announced. You’ll be told when the interview switches to system design vs. LLD vs. wrap-up.
• No live grading. The interviewer stays neutral in the moment. Feedback comes in the report at the end.
• You can stop any time. Say “stop” or “wrap it up” and you’ll get a partial report scoped to what was covered.

Structure

cornerman/
├── SKILL.md                            # Orchestrator: intake → phase routing → report
├── mcp/
│   ├── server.py                       # Cornerman MCP: sessions, timers, problem picker
│   └── data/
│       ├── blind75.json                # 71 Blind 75 problems (id, category, difficulty, url)
│       ├── blind75-source.md           # Original repo README (source of truth)
│       └── build_blind75.py            # Regenerates blind75.json from the source
├── references/
│   ├── 00-intake-analyst.md            # Session start, resume + JD, seniority classification
│   ├── 01-experience-interviewer.md    # Project walk + impact probe + technical fallback
│   ├── 02-jd-alignment.md              # Bridge past experience to the target role
│   ├── 03-architect-hld.md             # System design (HLD, async take-home)
│   ├── 04-architect-lld.md             # Low-level design (live)
│   ├── 05-evaluator.md                 # Scoring rubric + final report generator
│   ├── 06-coding-interviewer.md        # Coding (async take-home + walkthrough)
│   ├── seniority-calibration.md        # Difficulty matrix used by every persona
│   └── question-banks/
│       ├── backend.md
│       ├── coding-blind75.md
│       ├── ml-llm.md
│       └── system-design.md
├── assets/
│   ├── logo.svg
│   └── report-template.md              # Fixed structure for the final feedback report
└── README.md

What each layer does

• Skill (SKILL.md + references/) — the interview logic. Works standalone as a skill; degrades gracefully if the MCP isn’t installed.
• MCP (mcp/server.py) — persistent state. Sessions, per-round timers, rubric scores, past-session lookup, Blind 75 problem picker. SQLite at ~/.cornerman/cornerman.db.

Prerequisites

• A Claude environment that supports skills (claude.ai, Claude Code, or Cowork).
• uv on your PATH — the MCP is a Python script that runs via uv run. Install with brew install uv on macOS or curl -LsSf https://astral.sh/uv/install.sh | sh.
• Optional but recommended: the scheduled-tasks MCP, for auto-grading async take-home rounds. Without it, timers still record but the timeout callback doesn’t fire.

Installation

Two steps. Install the skill folder, then register the MCP server.

1. Install the skill:

rm -rf ~/.claude/skills/cornerman
rsync -a --exclude='.git' --exclude='__pycache__' /path/to/cornerman/ ~/.claude/skills/cornerman/

2. Register the MCP server (Claude Code shown; adapt to your client):

claude mcp add cornerman --scope user -- uv run /path/to/cornerman/mcp/server.py

--scope user makes it available in every project. Drop it if you only want it in one project.

3. (Optional) install the scheduled-tasks MCP for auto-grading async take-home rounds. Without it, timers still record but the timeout callback doesn’t fire — you just have to come back and paste your submission manually.

Restart your Claude client after installing. Verify with claude mcp list — you should see cornerman connected.

Skill-only mode

If you don’t have uv or don’t want the MCP, step 1 alone works. You’ll lose persistent state, past-session trend commentary, the automatic Blind 75 picker (falls back to references/question-banks/coding-blind75.md), and scheduled auto-grading. You keep: all the interview logic and the report.

Rebuilding the Blind 75 catalog

The catalog is derived from jaimin-bariya/blind-75-leetcode. To refresh:

cd cornerman/mcp/data && python3 build_blind75.py

Domain coverage

Question banks are seeded for backend, ML/LLM, general SWE / system design, and coding (Blind 75, 71 problems, filtered by tier). Cornerman follows your actual project domain — the banks are inspiration for most rounds; the Blind 75 is used directly for coding (via mcp__cornerman__pick_problem).

How it works under the hood

Three moving parts cooperate:

1. The skill loads persona files phase by phase. Each persona adopts a specific voice (analyst, interviewer, architect, evaluator), runs its round, then hands off.
2. The cornerman MCP persists everything to a SQLite file at ~/.cornerman/cornerman.db: sessions, dossier state, per-round start/end timestamps, and rubric scores. It also serves the Blind 75 catalog for the coding round.
3. The external scheduled-tasks MCP (optional) is called for take-home timeouts. When you say “ready” on a coding or HLD round, Cornerman calls round_start (records start_iso, returns auto_grade_at_iso), then creates a scheduled task that fires at auto_grade_at_iso with a self-contained callback prompt. If you submit before the timer, Cornerman deletes the scheduled task. If you don’t, the task fires in a new conversation, loads your session, and grades what’s there.

Cornerman degrades gracefully at each layer: no scheduled-tasks → timers become best-effort in one conversation. No cornerman MCP → dossier lives in conversation memory only, no cross-session history.

Design notes

• Persona-per-file. Each “interviewer” is a self-contained reference file the orchestrator loads at the start of its phase. Different phase, different voice.
• Turn-by-turn is enforced. The orchestrator’s top rule is “ask one question, stop, wait”. Every persona file reinforces it.
• State is persistent when the MCP is installed. Dossier, timers, and scores go to SQLite so scheduled callbacks can resume state and past sessions can be recalled for trend commentary.
• Auto-detected seniority, with confirmation. The intake analyst reads your resume and shows you a calibration summary before the interview starts. Correct it if it’s off — misclassification here poisons every later phase.
• Audio comes from the client. Cornerman itself is text-in / text-out. If your client has voice mode (claude.ai voice, mobile), it works the same way through voice.

---

🥊 Get in the ring.