🐹 gophers

26 production-grade Go skills for Claude Code, Gemini CLI, and opencode. Battle-tested patterns from the Go community — codified as triggerable AI skills.

License: MIT Go Version Skills Claude Code Gemini CLI opencode PRs Welcome

Quick StartAgent PluginsSkills CatalogHow It WorksExamplesFAQ


🎯 Why gophers?

Most AI assistants write Go like a senior JavaScript engineer pretending to like semicolons. gophers plugs in 26 opinionated skills that teach Claude (and friends) to write Go the way the standard library does — small interfaces, errors as values, no magic.

"The bigger the interface, the weaker the abstraction." — Rob Pike Now your AI knows that, before it writes a 12-method UserManagerService.

What you get:

  • 🧠 Discoverable — Each skill has an explicit trigger description; Claude knows when to use it.
  • 📚 Progressive disclosureSKILL.md is ≤ 200 lines; deep dives live in references/.
  • 🪞 Opinionated — Every rule cites a community source (Effective Go, Google Style Guide, Uber, Rob Pike talks).
  • 🔌 Multi-platform — Same skill files run in Claude Code, Gemini CLI, and opencode.
  • Verifiable — Every skill ends with a checklist your AI can self-grade against.

🚀 Quick Start

Claude Code (via marketplace)

# Add the marketplace
/plugin marketplace add muratmirgun/gophers

# Install the plugin
/plugin install gophers@gophers

Claude Code (manual)

cd ~/.claude/plugins
git clone https://github.com/muratmirgun/gophers

Then restart Claude Code. Skills auto-load from skills/.

Gemini CLI

gemini extensions install https://github.com/muratmirgun/gophers

opencode

opencode plugin add github.com/muratmirgun/gophers

Manual / any other agent

git clone https://github.com/muratmirgun/gophers ~/.config/ai/gophers
# Point your agent's CLAUDE.md / AGENTS.md / GEMINI.md at the skills/ directory.

🔌 Portable Agent Plugins

gophers is also packaged as an Agent Plugins 1.0.0 plugin. The root plugin.json is the portable manifest. Compatible clients discover each immediate skill directory under root skills/ automatically.

Agent Plugins defines package contents and discovery. Each client controls installation, distribution, enablement, updates, marketplace publication, and user interface. The standard does not define a universal installation command. See the current compatible clients for client support.

Package part Portability
plugin.json Portable Agent Plugins manifest
skills/ Portable Agent Skills and canonical content
.claude-plugin/ Claude Code manifest and generated invocation compatibility
gemini-extension.json Gemini CLI integration
opencode.json OpenCode integration
agents/ Client-specific prompts; Agent Plugins 1.0.0 has no portable mapping

The Claude Code compatibility tree is generated from root skills. It retains user-invocable controls and existing OpenClaw metadata without adding non-portable fields to canonical skills. The package needs no mcp.json because its capabilities are instructions and reference files, not runtime MCP tools.

Maintainers can validate the complete package with one command:

scripts/validate.sh

This command validates the live Agent Plugins schema, all 26 skills through the pinned official skills-ref library, local links, client JSON files, generated files, and existing repository rules.


📦 Skills Catalog

26 skills, grouped by intent. In Claude Code, one is user-invokable (/go-code-review); the generated client layer hides the rest from the slash menu while keeping automatic activation.

🧱 Fundamentals — language mechanics, taught well

Skill Emoji Triggers when…
go-naming 🏷️ naming any identifier — packages, types, methods, errors
go-declarations 📝 declaring vars, consts, structs, maps, iota enums
go-control-flow 🔀 writing conditionals, loops, switches, type switches
go-functions ƒ organising functions in a file, designing signatures
go-data-structures 📊 choosing/operating on slices, maps, arrays, strings
go-packages 📦 creating packages, organising imports, structuring projects
go-error-handling ⚠️ writing, wrapping, inspecting, or logging errors
go-interfaces 🔌 defining/implementing interfaces, embedding, receivers
go-generics 🧬 deciding whether to introduce generics, writing constraints
go-functional-options ⚙️ designing constructors with 3+ optional parameters
go-defensive 🛡️ hardening API boundaries — copy, defer, time, panic discipline
go-code-style writing/reviewing for clarity, formatting, design priority

⚙️ Concurrency — goroutines without surprises

Skill Emoji Triggers when…
go-context 📦 designing context.Context flow, deadlines, request values
go-concurrency 🚦 writing goroutines, channels, select, mutexes, errgroup

🌐 Web & APIs — framework-agnostic delivery layers

Skill Emoji Triggers when…
go-clean-architecture 🏛️ scaffolding a service into Domain/Usecase/Repository/Delivery
go-grpc 📡 implementing or reviewing gRPC servers/clients
go-graphql 🌐 building a GraphQL API (gqlgen or graph-gophers)
go-swagger 📋 adding OpenAPI/Swagger annotations with swaggo/swag

🗄️ Data & Observability — production signal

Skill Emoji Triggers when…
go-database 🗄️ writing SQL access code — sqlx/sqlc/pgx/GORM trade-offs
go-logging 📝 choosing a logger, configuring slog, request-scoped fields
go-observability 📈 instrumenting metrics, traces, exemplars, correlation
go-performance profiling, benchmarking, optimising — pprof decision tree

🧪 Quality & Process — ship safely

Skill Emoji Triggers when…
go-testing 🧪 writing tests — table-driven, subtests, fuzz, synctest, goleak
go-linting 🧹 setting up golangci-lint, suppressing findings, CI gates
go-documentation 📚 writing godoc comments, Example tests, README/CHANGELOG
go-code-review 👀 user-invokable/go-code-review walks a diff topic by topic

🔬 How It Works

Each skill is a single markdown file (SKILL.md) with structured frontmatter and a strict body shape:

---
name: go-interfaces
description: Use when defining or implementing Go interfaces...   # ← trigger
license: MIT
metadata:
  author: muratmirgun
  version: "0.1.0"
allowed-tools: Read Edit Write Glob Grep Bash(go:*)
---

# Title
1-2 sentence philosophy.

## Core Rules            ← 5-7 numbered, non-negotiable invariants
## Decision Table        ← when to apply / when not to
## Body sections         ← code examples, contrasts (Good / Bad)
## Anti-Patterns         ← table of common mistakes + fixes
## Verification Checklist← AI self-grades before claiming done
## References            ← links to deeper references/*.md

When Claude (or Gemini / opencode) reads code that matches the trigger, the skill is injected into context — opinionated rules + code examples + a verification checklist. Your AI assistant goes from "knows Go" to "writes Go like a stdlib author".


💡 Examples

Before gophers

type UserManagerInterface interface {
    GetUser(id string) (*User, error)
    SetUser(u *User) error
    DeleteUser(id string) error
    ListUsers() ([]*User, error)
    CountUsers() (int, error)
}

func GetUser(id string) (*User, error) {
    user, err := db.QueryUser(id)
    if err != nil {
        return nil, fmt.Errorf("db error: " + err.Error())
    }
    return user, nil
}

After gophers (go-interfaces + go-error-handling + go-naming fire)

// Reader fetches a User by ID. Returns ErrNotFound when absent.
type Reader interface {
    User(ctx context.Context, id string) (*User, error)
}

func (s *Store) User(ctx context.Context, id string) (*User, error) {
    u, err := s.db.User(ctx, id)
    if err != nil {
        return nil, fmt.Errorf("store: user %s: %w", id, err)
    }
    return u, nil
}

What changed:

  • 5-method UserManagerInterface → 1-method Reader (small interfaces compose)
  • GetUserUser (Go style: no Get prefix)
  • "db error: " + err.Error()%w (preserves errors.Is / errors.As)
  • context.Context first param (cancellation propagates)

🎨 The gophers Philosophy

Tenet What it means in practice
Errors are values No panic-as-exception, no swallowed errors, wrap with %w
Accept interfaces, return concrete types Consumers state needs; producers expose what they have
The framework is a detail Gin/Echo/Fiber lives in internal/delivery/, nothing else
The database is a detail SQL lives in internal/repository/, nothing else
Tests fail usefully Function(input) = got, want want — always
Documentation is part of the API godoc renders in IDE tooltips; signature noise is wasted ink
Measure before optimising pprof first, intuition last
Don't design with interfaces — discover them Wait for the second implementation

🛠️ Project Structure

gophers/
├── plugin.json                  # Portable Agent Plugins 1.0.0 manifest
├── .claude-plugin/
│   ├── plugin.json           # Claude Code plugin manifest
│   ├── marketplace.json      # Claude Code marketplace listing
│   ├── skill-overrides.json  # Client-only invocation and OpenClaw values
│   └── skills/               # Generated Claude/OpenClaw compatibility files
├── .github/workflows/
│   └── validate.yml          # Complete package validation
├── gemini-extension.json     # Gemini CLI extension manifest
├── opencode.json             # opencode plugin manifest
├── skills/                   # 26 canonical portable skills
│   └── go-<name>/
│       ├── SKILL.md          # ≤ 200 lines, opinionated rules
│       └── references/       # Deep dives, examples, cheat-sheets
├── agents/                   # Subagent prompts (extensible)
├── scripts/                  # Generation, tests, and validation
├── CLAUDE.md                 # Project context for AI assistants
└── README.md                 # You are here

❓ FAQ

No. Each skill activates independently based on its trigger description. If you never write GraphQL, go-graphql never fires. The cost of an unused skill is zero tokens.

Yes. The skills are plain markdown — usable as system prompts for any LLM. The plugin manifests just automate discovery for Claude Code, Gemini CLI, and opencode.

Token budget. A 5,000-line style guide poisons context. 26 focused skills with explicit triggers load only what's relevant to the current diff.

Yes — go-linting ships an opinionated .golangci.yml and the other skills cite the same checks. No conflicts.

Go 1.21+ baseline. A few skills reference Go 1.24+ (b.Loop) and Go 1.25+ (testing/synctest) — they call out the version explicitly.

Open an issue with the skill name, the trigger conditions, and 2-3 concrete rules it would enforce. We reject vague "best practices" skills — every skill must have a verifiable checklist.


🤝 Contributing

PRs welcome — but the skill bar is high:

  1. Trigger must be unambiguous. "Use when X" — not "Use when working on Go".
  2. Every rule cites a source. Effective Go, Google Style Guide, Uber, a standard-library API, or a Rob Pike talk. No bare opinions.
  3. SKILL.md ≤ 200 lines. Deep content goes in references/.
  4. Every skill ends with a verification checklist. Items must be observable (a go vet flag, an errors.Is call, a grep pattern).
  5. No emoji in body text unless the user requested them. Frontmatter emoji: field is the only exception.

See CLAUDE.md for the full authoring checklist.


📜 License

MIT © muratmirgun

Influenced by:


⬆ back to top

Built with Claude Code. Reviewed by Claude Code. Used by Claude Code.