What it is
Six Agent Skills act as a system design reference and assistant on real projects. They scan a codebase for the stores, queues, caches and transports it runs, check them against a review checklist, estimate the next 10× with a calculator, and write design documents and review recommendations with trade-offs.
The knowledge comes from all 28 case studies of Alex Xu’s System Design Interview – An Insider’s Guide and its Volume 2 with Sahn Lam, as summarised in the system-design-notes repository. The notes carry no licence and the books are copyrighted, so each reference is a rewrite, not a quotation, and the figures are left out. Interview practice is a mode, not the point.
The skills work in any tool that supports the Agent Skills open format, so Claude Code, Cursor and Codex all run them.
Install
npx skills add https://github.com/alexey1312/System-Design-Agent-Skill
One skill instead of all six:
npx skills add https://github.com/alexey1312/System-Design-Agent-Skill --skill system-design-transactions
Claude Code takes the plugin instead:
/plugin marketplace add alexey1312/System-Design-Agent-Skill
/plugin install system-design-skills@system-design-skills
Then ask a real question, for example:
Review this repo for design problems before we grow 10x. Don’t change anything.
Should we put Redis in front of the orders table or add read replicas?
Two users keep booking the last room. Fix the data model and the locking.
The six skills
| Skill | What it covers |
|---|---|
system-design-advisor | Orchestrator: scan, numbers, chapter map, options, recommendation, rollout; design documents, reviews, capacity plans, interview practice |
system-design-infrastructure | Rate limiter, consistent hashing, key-value store, unique IDs, message queue, object storage, metrics |
system-design-content-platforms | URL shortener, web crawler, news feed, autocomplete, video streaming, cloud file storage |
system-design-realtime-messaging | Notifications, chat, email, gaming leaderboard, ad click aggregation |
system-design-geospatial | Proximity service, nearby friends, maps and navigation |
system-design-transactions | Hotel reservation, payment system, digital wallet, stock exchange |
Context before boxes
A design answer that starts with a diagram is a guess. The skills treat “design X” or “review this” as permission to read and ask, not to draw. Before any architecture, you get:
the inventory of what the project already runs, with evidence;
the requirements, the targets and each assumption, written where you can challenge it;
an estimate table with its derivation, from metrics or the calculator;
the chapter behind each pattern, or the label inference.
Then come the options, the recommendation, the deep dives that carry the risk, and a section on rollout and verification. Each recommendation has a stable ID and a status: proposed, accepted, declined, kept or needs-data. A documented, deliberate choice that contradicts a chapter is reported as kept, not rewritten.
How it reads a project
Two read-only Python scripts do the measuring, so the agent reasons over structured output instead of ad-hoc grep.
arch_scan.py reads manifests, Docker Compose, Kubernetes, Terraform and imports. It reports the databases, caches, queues, search indexes, storage, push, payment and geo libraries a project uses, each with file:line evidence. It also flags the patterns the books warn about: money in floats, hash % N placement, SELECT … FOR UPDATE, timestamp-based optimistic locks and rank by COUNT(*).
estimate.py turns daily users, actions and payload sizes into QPS, storage, bandwidth, cache size and a server count, and shows the derivation of each number. A review checklist of 45 checks in nine groups cites a chapter for each check.
Evaluation
Four evals ran once with the skills and once without. With the skills, each of the 25 expectations passed; without them, 59 % passed. The model alone found the same bugs in the fixture project. What the skills add is what a reference is for: a chapter for each pattern, a scan before any recommendation, recommendation IDs with statuses, and the calculator’s tables in place of arithmetic in prose. They cost about a third more tokens per run.