swe-interview-coach

Turn Claude Code into your SWE interview coach — behavioral, system design & coding mocks, all local.

Get started in 10 seconds

/plugin marketplace add kirilxd/swe-interview-coach
/plugin install swe-interview-coach@kirilxd-plugins

Then open Claude Code in your interview-prep directory (e.g., ~/Documents/interview/) and run:

/story-add ~/Documents/your-resume.pdf

What it does

Builds a canonical bank of STAR stories from your résumé or experience, maps them to each company you're targeting (with their values and JD), runs realistic mock interviews, drills story delivery, and captures learnings from real interviews. Everything is plain markdown in your working directory — you can edit anything in vim and the plugin still works.

Interview modes

The data layout under ~/Documents/interview/ is designed for these modes from day one — adding a take-home/ subtree later requires no migration.

Behavioral commands

Sysdesign commands

How the canvas works

The first sysdesign command you run boots a tiny bundled Node-stdlib server (scripts/canvas-server.js) on http://localhost:9999 and opens a browser tab. Excalidraw loads once from a pinned CDN (esm.sh) and is then browser-cached — only the first load needs internet. The scene syncs both directions through one local JSON file, ~/.config/swe-interview-coach/canvas.json: the agent's writes appear in your browser within ~500 ms, and your edits post back within ~300 ms. Roles shift per mode — explain: the agent draws; practice: the agent scribes and you can edit alongside it; mock: you draw while the agent observes, then annotates afterward. Prerequisites are just node, lsof, and a browser. There is no MCP, no npm install, and no /reload-plugins — the script is plain Node standard library.

Coding commands

How the test harness works

A practice or mock session seeds a solution.py in the session folder; you edit it in your own editor (vim, VS Code, whatever). On a mock, the coach runs your solution against the problem's bundled test cases through a tiny bundled Python-stdlib runner (scripts/run-solution.sh → scripts/harness/python_runner.py) and scores correctness from the result. The only prerequisite is python3 on your PATH; if it isn't available the harness degrades gracefully to a read-only review of your code — coaching never aborts over a missing runtime. There is no MCP, no npm install, and no server — it's plain Python standard library that runs entirely on your machine.

The behavioral prep loop

build → map → mock → drill → debrief → repeat

1. Build your bank — /story-add (with résumé) or /story-import (from existing notes). Each story can have multiple framings ("variants") tuned to different question themes.
2. Map for the target — /story-map <company_name> <jd-url> researches 's values, fetches the JD, tailors every story, and builds a question index with coverage gaps.
3. Mock it — /mock-behavioral <company_name> runs an interview as a engineer. Asks 5–7 questions with follow-ups. Scores STAR clarity, metric quality, conciseness, and values alignment on a 1–5 rubric.
4. Drill the weak spots — /story-rehearse <id> <variant> for delivery polish. The mock's coach notes tell you exactly which variants to drill.
5. Debrief after real interviews — /debrief <company_name> captures the actual questions asked, what stories you told, what surprised you. New questions feed back into the bank for next time.

The sysdesign prep loop

explain → practice → mock → debrief → repeat

1. Learn the pattern — /sysdesign-explain twitter-feed walks you through a canonical design senior-engineer style, drawing the architecture on the canvas as it teaches.
2. Practice untimed — /practice-sysdesign url-shortener lets you talk through an architecture while the agent scribes your boxes and arrows and hints freely whenever you're stuck.
3. Mock under pressure — /mock-sysdesign uber-dispatch runs a graded 45-minute mock: you drive the diagram, the agent probes, then scores a 6-axis rubric and annotates your canvas.
4. Debrief after real interviews — /debrief-sysdesign <company_name> rate-limiter captures the actual problem, your approach, where you got stuck, and your gut on the outcome.

The coding prep loop

explain → practice → mock → drill → debrief → repeat

1. Learn the pattern — /coding-explain sliding-window walks you through a DSA pattern senior-engineer style, worked on a representative problem so you internalize the shape, not just one answer.
2. Practice untimed — /practice-coding two-sum seeds a solution.py you edit in your own editor while the coach guides you and hints freely whenever you're stuck.
3. Mock under pressure — /mock-coding coin-change runs a graded mock: you code in solution.py, the coach probes, then runs your code against the bundled cases and scores correctness, complexity, and communication.
4. Drill the weak spots — /coding-drill dynamic-programming runs rapid-fire reps on a pattern; with no argument it targets the patterns your mocks flagged as weak.
5. Debrief after real interviews — /debrief-coding <company_name> two-sum captures the actual problem, your approach, where you got stuck, and your gut on the outcome.

Your data stays local

All state lives in your working directory (typically ~/Documents/interview/). No hidden home directory, no JSON, no database.

~/Documents/interview/
├── behavioral/
│   ├── stories/                          # canonical STAR bank (cross-company)
│   │   └── <company_name>-ui-library.md
│   └── rehearsals/                       # canonical drill sessions
├── sysdesign/
│   └── sessions/                         # cross-company practice & mock runs
│       ├── 2026-06-13-mock-sysdesign/    # transcript.md + canvas.excalidraw + canvas-annotated.excalidraw
│       └── 2026-06-13-practice-sysdesign/ # transcript.md + canvas.excalidraw
├── coding/
│   ├── sessions/                         # cross-company practice & mock runs
│   │   ├── 2026-06-13-mock-coding/       # solution.py + transcript.md (with rubric)
│   │   └── 2026-06-13-practice-coding/   # solution.py + transcript.md
│   ├── drills/                           # rapid-fire pattern reps
│   │   └── 2026-06-13-drill-dynamic-programming.md
│   ├── imported/                         # /coding-import'ed LeetCode problems
│   │   └── two-sum.md
│   └── weak-patterns.md                  # rolling tracker of flagged patterns
├── <company_name>/
│   ├── values.md                         # cached company culture/values
│   ├── jds/
│   │   └── senior-fe-engineer.md         # JDs you've prepped for
│   ├── behavioral/
│   │   ├── mapped-stories/
│   │   │   └── senior-fe-engineer/       # company-tailored variants
│   │   └── sessions/                     # mocks, mapped rehearsals, debriefs
│   ├── sysdesign/
│   │   └── sessions/
│   │       └── 2026-06-13-debrief-sysdesign.md
│   └── coding/
│       └── sessions/
│           └── 2026-06-13-debrief-coding.md
└── <company_name>/
    └── (same per-company structure)

Markdown with YAML frontmatter throughout. The coding/ mode now sits alongside behavioral/ and sysdesign/ under each company directory the same way; take-home/ will slot in later with no migration.

Privacy

• Everything stays on your machine. No data is sent to external services beyond the standard Claude Code conversation.
• Company values caches use WebSearch and WebFetch for the same company-facing pages a human would Google.
• Your résumé is read locally via Claude's Read tool; only excerpts you discuss are processed.
• The sysdesign canvas runs entirely on localhost:9999 via a bundled stdlib Node script; scene state lives in ~/.config/swe-interview-coach/canvas.json and never leaves your machine. The Excalidraw library is fetched once from a pinned esm.sh CDN and then cached, and canvas exports save into your working dir's session folders.
• The coding test harness runs your code locally via a bundled Python stdlib script; nothing is sent anywhere. It executes code without a sandbox (full local filesystem/network access), so only run solutions you trust — this is your own code on your own machine. /coding-import is the only coding command that reaches the network, doing a read-only public fetch (WebFetch) of a LeetCode problem page — no login, no submission.

Local development

claude --plugin-dir ~/Documents/swe-interview-coach

After editing any command, agent, or skill file, run /reload-plugins in your session.

Uninstalling

/plugin uninstall swe-interview-coach@kirilxd-plugins
/plugin marketplace remove kirilxd-plugins

To remove your interview data: rm -rf ~/Documents/interview/ (and any per-company subdirs you've created).

License

MIT