code-design-patterns

code-design-patterns is a free, open-source Claude Skill for software design patterns. It is a folder of Markdown — one SKILL.md router plus eleven reference files — that Claude loads automatically when the work involves class design, refactoring, architecture or a low-level design problem. It covers 115 pattern entries, 21 anti-patterns, a 10-dimension evaluation rubric and 30 canonical LLD problems, and it is MIT licensed.

Clone the repo and copy one folder into your skills directory — there is no build step and no runtime dependency:

git clone https://github.com/00suryavanshi00/code-design-patterns.git
cp -r code-design-patterns/skill/code-design-patterns ~/.claude/skills/

Use .claude/skills/ inside a repo for project scope, or ~/.claude/skills/ for every project. For Claude.ai and Cowork, upload the folder under Settings → Capabilities → Skills. A packaged single-file .skill bundle is also available in the repo’s dist/ folder.

Claude already knows the catalogue. What it does not do reliably is decline to use it — so designs arrive with a Singleton nobody asked for, a Factory wrapping new, and an Observer with no subscribers, while the race condition in the allocation path goes unmentioned. This skill inverts the default. Every pattern in the output must trace to a specific, stated force in the requirements; if the force cannot be named, the pattern is removed before you see it. Its main job is subtraction, not addition.

All 23 Gang of Four patterns — Strategy, Factory Method, Abstract Factory, Builder, Prototype, Singleton, Adapter, Bridge, Composite, Decorator, Facade, Flyweight, Proxy, Chain of Responsibility, Command, Interpreter, Iterator, Mediator, Memento, Observer, State, Template Method and Visitor — plus architectural patterns (dependency injection, repository, hexagonal / ports-and-adapters, CQRS, event sourcing, result types, plugin systems, feature flags), concurrency patterns (worker pool, pipeline, fan-out/fan-in, actor, single-writer, compare-and-swap, backpressure), distributed-resilience patterns (circuit breaker, retry with budget, hedged requests, bulkhead, saga, transactional outbox, idempotency key, five rate-limiting algorithms, expand-contract migration) and frontend component patterns (compound components, custom hooks, provider splitting, props getters, state reducers, error boundaries), persistence patterns (optimistic and pessimistic locking, isolation levels, write skew, unique constraints, conditional writes, aggregates as transaction boundaries, multi-tenant isolation) and API contract patterns (idempotent HTTP semantics, error contracts, pagination, compatible evolution, versioning, deprecation, webhooks). 115 entries in total, across eleven reference files.

Patterns are ranked by how often they are actually the right answer in production code, not by how famous they are — ★★ confident, ★ situational, · last resort. The idea is borrowed from Christopher Alexander’s A Pattern Language, the 1977 book the Gang of Four took the term “design pattern” from, which rated its 253 patterns the same way. Singleton sits in the bottom tier because in most codebases it is global mutable state with a nicer name, and passing the dependency explicitly is both simpler and testable.

Twenty-one code smells, each with a detectable signature and the smallest first refactoring move — including pattern-stuffing, god object, anemic domain model, primitive obsession, feature envy, shotgun surgery, leaky abstraction, singleton abuse, premature abstraction, boolean-parameter explosion, check-then-act races, and writing to a database and a queue as two separate commits. Pattern-stuffing is the one the whole skill exists to prevent.

Yes. It ships a bank of 30 canonical low-level design problems — parking lot, elevator system, vending machine, rate limiter, LRU cache, Splitwise, chess, ATM, URL shortener, message queue and more. Each entry carries the axis of change, the patterns that legitimately fit, the plausible wrong answer, and the follow-up probe that breaks a memorised one. You can use it to practise, to generate problems, or as ground truth when grading a design out of 30.

It is built and tested for Claude — Claude Code, Claude Desktop, Claude.ai and Cowork. Because it is plain Markdown in the SKILL.md convention with zero dependencies and no runtime, any agent that reads SKILL.md folders can load it as-is; the routing front-matter is the only Claude-specific part, and it degrades to an ordinary heading elsewhere.

Yes — MIT licensed, free for personal and commercial use, no account, no telemetry, no network calls. The source is on GitHub.

Before a design is shown to you, Claude scores it against ten dimensions — each 0–3, for a total out of 30 — covering requirement traceability, the named axis of change, cohesion and coupling, concurrency correctness, failure handling, testability, observability, cost, simplicity, and whether every pattern used was justified. Automatic red flags, such as a pattern with no named force or a Manager class, cap the total at 20 no matter how the rest scores.