Repo of the day: claude-code-workflows, from requirement to commit with specialized agents

💡 What it is: one marketplace, four plugins

The repository by shinpr is not a single plugin but a marketplace for Claude Code: you add it once and install the piece you need. dev-workflows covers backend, APIs and CLIs; dev-workflows-frontend adds React/TypeScript agents; dev-workflows-fullstack merges the two worlds into a single plugin; dev-skills is the minimal variant, rules and best practices only, for those who already have their own orchestration.

The promise is concrete: the code produced matches its design docs and passes its tests, because every phase — requirements, design, implementation, quality checks — is handled by a dedicated agent working in a fresh context. Phases don't pollute each other, which is the typical flaw of long single-agent sessions.

# 1. add the marketplace
/plugin marketplace add shinpr/claude-code-workflows

# 2. install the plugin for your stack
/plugin install dev-workflows@claude-code-workflows
/reload-plugins

# 3. build a feature end-to-end
/recipe-implement "user authentication with JWT"

🧭 How it works: work gets routed by size

The heart of the system is the requirement-analyzer: it evaluates how big the request is and picks the path. A small task (1-2 files) goes straight to implementation; a medium one passes through codebase analysis; a large one (6+ files) traverses the whole pipeline with PRD, design docs, document reviews and a work plan split into committable tasks.

It's the smartest part of the project: no bureaucracy on two-line fixes, no improvisation on big features. Ceremony is proportional to risk, decided by an agent rather than by the rush of the moment.

🤖 Specialized agents: a team, not a jack-of-all-trades

Under the hood there are more than twenty agents with precise roles. Some are shared across plugins, others are backend- or frontend-specific. The division of roles mirrors a real team: whoever analyzes doesn't implement, whoever implements doesn't promote themselves to reviewer.

Three examples give a sense of the level of detail the project has reached:

• task-executor implements with TDD and produces commit-ready chunks, while quality-fixer runs tests, type checks and lint and fixes what it finds.
• investigator, verifier and solver form the diagnosis flow: they map execution paths, validate failure points with the Devil's Advocate method and propose solutions with trade-off analysis.
• security-reviewer combs through the implementation at the end, and design-sync verifies consistency across multiple design docs — precious on fullstack work, where backend and frontend have separate documents.

🍳 Recipes: one command per scenario

The user interface of the system are the recipes, slash commands with a common prefix: type the slash and autocompletion shows every available flow. The three we use most in our tests: recipe-implement for the full feature, recipe-diagnose to investigate a bug before touching code, recipe-task for the surgical change with quality checks.

The most surprising one is recipe-reverse-engineer: it starts from existing code and generates PRDs and design docs verified against the real implementation. For anyone with an undocumented legacy codebase it's a serious accelerator: it makes the codebase readable to AI agents even before people.

/recipe-implement "new feature end-to-end"
/recipe-task "targeted fix with quality checks"
/recipe-diagnose "API returns 500 on login"
/recipe-reverse-engineer "src/auth"

# fullstack (dedicated plugin)
/recipe-fullstack-implement "JWT + login form"

⚖️ With or without orchestration: dev-workflows vs dev-skills

A design choice we appreciate: the project doesn't force its orchestration on you. If you already have your own prompts, scripts or a CI-driven loop, you install dev-skills and take only the rules — coding principles, testing principles, documentation standards — which agents load when relevant.

Mind one detail the README documents well: the plugins share the same skills, so they must not be installed together. Claude Code reserves about 2% of the context window for skill descriptions: exceeding it makes skills silently ignored.

🏢 Why it matters for developers and SMEs

For developers the value is immediate: battle-tested orchestration patterns to study or borrow — fresh contexts per phase, routing by complexity, quality gates before commit. They are the same ideas we apply in our structured vibe coding projects, and finding them packaged in an MIT plugin shortens the curve for anyone starting today.

For an SME the point is different: the gap between «we tried AI» and «AI is in the process» sits exactly here. A flow where every feature is born with written requirements, design docs verified against code and passing tests is an auditable flow, where the agent's work can be controlled like a supplier's. The ecosystem around it — metronome against agent shortcut-taking, pr-review for GitHub PR reviews, the codex-workflows twin for Codex CLI users — shows the project is alive: last push on July 6, version 0.22.x, over 140 merged PRs.

repo      shinpr/claude-code-workflows
stars     ~500 · MIT license
activity  last push 2026-07-06 · v0.22.3
contents  4 plugins · 20+ agents · recipe-*
extras    metronome · pr-review · codex-workflows

🗺️ Where to start

The path we suggest: add the marketplace, install dev-workflows on a secondary project and run a recipe-implement on a real but non-critical feature. Look at the documents it generates in docs/prd and docs/design before even reading the code: that's where you understand whether the method is for you. Remember to ignore docs/plans in your gitignore — it's ephemeral working state, not documentation.

If your team wants to bring this kind of discipline into its development process — with Claude Code, Codex or custom tooling — that's exactly the ground we work on every day: talk to us and let's build the workflow that fits your context.

More articles from the blog