Mrbaeksang

Onboard any new external dependency before using it — research the official docs, vendor them as the single source of truth, then generate a blocking guard. Use when adding or first using any library, framework, SDK, API, or CLI tool; when editing package.json / requirements.txt / pyproject.toml / go.mod / Cargo.toml; or when the user mentions adopting a new tool. This is a rail, not advice — the trigger is unconditional, never "when uncertain".

Configure the mrbaeksang rails in a repo — record its deployment level and confirm the enforcing hooks are active. Run this once after installing the plugin. Use when the user runs /setup-mrbaeksang-skills or asks to set up / configure the rails.

The rigor gate that refuses LLM avoidance. Use at the start of any non-trivial task, and the instant you're about to say "that's overkill", "we're blocked on X", "we'd need a credential / DB / API key", "I'm not sure", or about to hand-author a spec/schema/API doc. Calibrates rigor to the repo's deployment level, forces a test-double or local instance before claiming blocked, research before asserting, and chains onboard-dependency for any new dependency. One gate; it chains the rest.

Enforces Karpathy's 4 LLM-coding principles (Think Before Coding, Simplicity First, Surgical Changes, Goal-Driven Execution). Auto-fires whenever the user requests writing, editing, refactoring, fixing, or adding code. Blocks drive-by refactoring, over-engineering, and speculative changes; demands clarification before implementation, minimum code, surgical edits, and conversion of tasks into verifiable goals. Respond in the user's language.

Service-profile-driven project audit. Auto-fires when the user requests audit, review, code review, pre-launch check, security audit, OWASP/SOLID/12-Factor compliance, project skeleton/bootstrap/setup, or any equivalent in any language (e.g., 점검, 감사, 리뷰, 출시 전 검토, 보안 점검, 골조, 셋업). Reads the full 0–10 section checklist from SPEC.md, filters items by grade (🔴🟠🟡🔵⚪) against the user's service profile, and outputs results as ✅❌⚠️⏭️ markers re-sorted by risk. Respond in the user's language.

Deep research that reads the whole web, including pages behind a Bot Wall (Cloudflare, Datadome, Turnstile, reCAPTCHA). Use when the user asks for a researched, cited answer — especially on topics where sources are likely paywalled, rate-limited, or bot-protected, or when an ordinary search/research tool returned thin or blocked results.

Use when the user wants to work with Flow (flow.team) — reading or managing projects, tasks, schedules, todos, the calendar, team members, search, or alarms via the Flow Open API. Covers status checks ("what's on my plate", "any unread alarms", "this week's schedule"), creating and updating work (tasks, posts, schedules, calendar events), and team lookups. After reading state, proactively surface what needs attention (overdue tasks, unread alarms, today's events) and offer to act.