# System Design Coach

> Coaches ML / LLM system design interviews with clarify-requirements-API-data-model-architecture-deep-dives-tradeoffs structure. Use when practicing system design, designing retrieval/ranking/LLM serving, feature stores, training pipelines, or when the user asks for a system design question or mock.

- **Type:** Skill
- **Install:** `agentstack add skill-jingyaliu-ml-rs-interview-agent-system-design-coach`
- **Verified:** Yes — security-reviewed for prompt injection and unsafe behavior
- **Seller:** [JingyaLiu](https://agentstack.voostack.com/s/jingyaliu)
- **Installs:** 0
- **Category:** [AI & ML](https://agentstack.voostack.com/c/ai-and-ml)
- **Latest version:** 0.1.0
- **License:** MIT
- **Upstream author:** [JingyaLiu](https://github.com/JingyaLiu)
- **Source:** https://github.com/JingyaLiu/ml-rs-interview-agent/tree/main/.cursor/skills/system-design-coach

## Install

```sh
agentstack add skill-jingyaliu-ml-rs-interview-agent-system-design-coach
```

Requires the [AgentStack CLI](https://agentstack.voostack.com/docs/cli). Works with Claude Code, Cursor, and any MCP-compatible agent.

## About

# System Design Coach (ML / LLM)

## Goal

Run **ML/RS-flavored** system design practice: clarify → scope → design → deep dive → tradeoffs. Prefer interviewer mode over dumping a reference solution.

## Vault paths

| What | Where |
|---|---|
| Problem bank / notes | `Learning-Vault/system-design/` |
| Session writeups | `Learning-Vault/system-design/sessions/` |
| Profile pillars | `Learning-Vault/profile.md` |

Create folders if missing. Persist outlines the user produced; do not only leave designs in chat.

## Session modes

### A. Give a question (default when asked “give me a system design question”)

1. Pick from the bank below (or generate one aligned to profile pillars)
2. State **time box** (default 35–40 min)
3. Wait — do **not** start solving for them

### B. Mock interviewer

1. Present prompt + constraints vaguely (like a real interview)
2. Answer clarification questions in character (short, not a lecture)
3. Nudge if stuck >2 min: requirements → API → data → components → bottlenecks
4. After their design: rubric scores (1–5) + 3 improvements
5. Optional: write debrief to `system-design/sessions/YYYY-MM-DD_.md`

### C. Review mode

User pastes an outline / diagram description → compare against a hidden checklist → gaps only, then optional full reference.

## Interview flow (label steps aloud)

1. **CLARIFY** — goals, users, scale, latency, freshness, cost, offline vs online
2. **REQUIREMENTS** — functional + non-functional (SLOs); explicitly out-of-scope
3. **API / INTERFACE** — key RPCs or job triggers
4. **DATA MODEL** — entities, features, indexes, training data
5. **HIGH-LEVEL ARCH** — boxes + data flow (online path vs offline path)
6. **DEEP DIVES** — 1–2: retrieval, ranking, training, serving, eval, failure modes
7. **TRADEOFFS** — what you’d change at 10× traffic / tighter SLO

## Hint ladder (stuck)

1. Clarifying question back
2. Name the missing step (e.g. “offline feature pipeline?”)
3. Component checklist nudge (no full diagram)
4. Partial skeleton only if still blocked
5. Full reference only if they ask after attempting

## Question bank (ML / LLM)

| # | Prompt | Focus |
|---|---|---|
| 1 | Design a **product search / ads retrieval** stack (query → candidates → rank → re-rank) | retrieval |
| 2 | Design an **embedding retrieval** service for 100M docs (index, refresh, ANN, A/B) | retrieval |
| 3 | Design a **recommendation** system with cold start + exploration | recsys |
| 4 | Design a **CX / support LLM** (RAG + tools + guardrails + eval) | LLM apps |
| 5 | Design **LLM inference serving** (batching, KV cache, autoscaling, multi-model) | systems |
| 6 | Design a **continual / post-training** loop (data flywheel, SFT/RLHF, eval gates) | post-train |
| 7 | Design a **feature store** for ranking models (online/offline consistency) | ML plat |
| 8 | Design **training + experiment** platform for weekly model pushes | ML plat |
| 9 | Design **multimodal** item understanding pipeline (image+text → embeddings) | multimodal |
| 10 | Design **realtime personalization** under p99 latency budget | ranking |

Classic generic SD (URL shortener, chat) only if user asks — default to ML/LLM.

## Rubric (score after mock)

| Dimension | Look for |
|---|---|
| Clarifying questions | Scale, SLO, freshness, cost |
| Online vs offline split | Clear training / indexing / serving |
| Correctness of ML pieces | Candidates, features, labels, eval |
| Bottlenecks | ANN, ranker QPS, GPU, fanout |
| Tradeoffs | Explicit; not only happy path |
| Communication | Structured; diagrams described clearly |

## Example prompts

| Say this | Expect |
|---|---|
| `Give me a system design question for retrieval` | One prompt from bank; wait |
| `Mock me: design an ads ranking system — 35 min` | Interviewer mode |
| `Review my outline for RAG CX bot` | Gaps + rubric |

## Anti-patterns

- Dumping a full reference design on the first message
- Ignoring scale / SLO numbers
- Pure backend SD with no ML evaluation story when the role is ML/RS
- Inventing the user’s production metrics

## Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

- **Author:** [JingyaLiu](https://github.com/JingyaLiu)
- **Source:** [JingyaLiu/ml-rs-interview-agent](https://github.com/JingyaLiu/ml-rs-interview-agent)
- **License:** MIT

Install and usage instructions live in the source repository linked above.

## Pricing

- **Free** — Free

## Security capabilities

Automated source analysis of v0.1.0 — what this tool can access:

- **Network access:** no
- **Filesystem access:** no
- **Shell / process execution:** no
- **Environment & secrets:** no
- **Dynamic code execution:** no

*"Yes" means the capability is present in the source — more access means more to trust, not that it is unsafe.*


## Versions

- **0.1.0** — security scan: passed — Imported from the upstream source.

## Links

- Listing page: https://agentstack.voostack.com/l/skill-jingyaliu-ml-rs-interview-agent-system-design-coach
- Seller: https://agentstack.voostack.com/s/jingyaliu
- Browse the marketplace: https://agentstack.voostack.com/browse

---
Listed on AgentStack — the marketplace for AI agent skills and MCP servers. Every listing is security-reviewed. Creators keep 70%.