# Claude Code 7-Phase Feature Development Workflow: A Complete Guide > Automate full-stack feature development with Claude Code's 7 phase workflow. This AI driven process ensures features are understood, designed, and reviewed for robust implementation. - Repository: [Anthropic/claude-code](https://github.com/anthropics/claude-code) - Tags: how-to-guide - Published: 2026-04-02 --- **Claude Code's feature-dev plugin automates full-stack feature implementation through a deterministic seven-step process driven by specialized AI agents that ensure every feature is well-understood, thoughtfully designed, and rigorously reviewed.** The `feature-dev` plugin in the `anthropics/claude-code` repository transforms informal feature requests into disciplined engineering workflows. This 7-phase approach for feature development workflow automation reduces rework by enforcing structured discovery, architecture design, and quality assurance before any code merges into the repository. ## The Seven Phases of Feature Development Claude Code orchestrates each feature through a predictable sequence from initial intent to final documentation. Every phase is encoded in the command definition at [`plugins/feature-dev/commands/feature-dev.md`](https://github.com/anthropics/claude-code/blob/main/plugins/feature-dev/commands/feature-dev.md) and documented in [`plugins/feature-dev/README.md`](https://github.com/anthropics/claude-code/blob/main/plugins/feature-dev/README.md). ### Phase 1: Discovery The workflow opens with **intent capture**. The system clarifies the problem statement, identifies constraints and requirements, and summarizes its understanding for developer confirmation. This ensures the AI comprehends the business goal before examining implementation details. ### Phase 2: Codebase Exploration Once the intent is clear, Claude Code launches **2-3 `code-explorer` agents in parallel** to build a mental model of the surrounding code. Each agent investigates a different angle—similar features, high-level architecture, or UI patterns—and returns a curated list of **5-10 important files** to read. This multi-perspective analysis prevents architectural mismatches and surfaces relevant abstractions early. ### Phase 3: Clarifying Questions Before generating designs, the system **eliminates ambiguity** by reviewing explorer findings against the original request. It generates a structured list of missing details covering edge cases, error handling, integration points, and performance needs, then waits for developer answers. This checkpoint prevents expensive rework during implementation. ### Phase 4: Architecture Design With requirements confirmed, Claude Code spawns **2-3 `code-architect` agents** with distinct strategies—minimal change, clean architecture, or pragmatic balance. Each agent produces concrete implementation blueprints annotated with trade-offs. The system synthesizes these approaches, compares trade-offs, and recommends the best fit, asking the developer to pick an approach before proceeding. ### Phase 5: Implementation After explicit developer approval, the assistant reads all files highlighted in Phase 2 and **writes the feature code** adhering to the chosen architecture and project conventions. Throughout this phase, Claude Code continuously updates a Todo list to track progress and maintain visibility into remaining work. ### Phase 6: Quality Review The workflow enforces quality through **three `code-reviewer` agents running in parallel**, each focusing on a specific dimension: simplicity and DRY principles, bugs and functionality, and project conventions defined in [`CLAUDE.md`](https://github.com/anthropics/claude-code/blob/main/CLAUDE.md). The system consolidates findings, surfaces high-severity issues with confidence scores, and lets the developer decide whether to fix immediately, defer, or proceed. ### Phase 7: Summary The final phase **closes all todos** and produces documentation summarizing what was built, key architectural decisions, modified files, and next steps. This creates a concise hand-off note for future reference and completes the audit trail. ## How the Specialized Agents Cooperate The 7-phase workflow relies on three distinct agent types defined in `plugins/feature-dev/agents/`. Each agent has a specialized prompt and output format encoded in its markdown definition file. - **`code-explorer`**: Defined in [`plugins/feature-dev/agents/code-explorer.md`](https://github.com/anthropics/claude-code/blob/main/plugins/feature-dev/agents/code-explorer.md), these agents scan the repository, trace execution paths, and surface key abstractions and file locations relevant to the feature. - **`code-architect`**: Defined in [`plugins/feature-dev/agents/code-architect.md`](https://github.com/anthropics/claude-code/blob/main/plugins/feature-dev/agents/code-architect.md), these agents synthesize exploration insights into multiple design blueprints, each annotated with trade-offs and implementation strategies. - **`code-reviewer`**: Defined in [`plugins/feature-dev/agents/code-reviewer.md`](https://github.com/anthropics/claude-code/blob/main/plugins/feature-dev/agents/code-reviewer.md), these agents enforce the project's style guide and report confidence-scored issues across correctness, readability, and compliance dimensions. Together, they turn an informal request like "Add OAuth login" into a reproducible, multi-stage engineering process. ## Running the Feature Development Workflow Initiate the full 7-phase automation using the `/feature-dev` command: ```bash # Start a full-feature workflow (Claude Code will drive all 7 phases) /feature-dev Add rate limiting to API endpoints ``` During the session, the assistant will execute each phase sequentially: 1. **Discovery** – Ask clarifying questions like "What traffic volumes do you expect?" 2. **Exploration** – Run `code-explorer` agents returning files such as [`src/api/rateLimiter.ts`](https://github.com/anthropics/claude-code/blob/main/src/api/rateLimiter.ts) and `src/middleware/` 3. **Clarifying Questions** – List open items like "Should the limit be per-IP or per-user?" 4. **Architecture** – Propose a "middleware-only" approach versus a "global service" approach 5. **Implementation** – After approval, write the new middleware, update tests, and commit 6. **Quality Review** – `code-reviewer` points out missing error-handlers and suggests DRY-friendly helpers 7. **Summary** – Post a concise report of changes and next steps For manual exploration outside the full workflow, launch individual agents: ```bash # Launch a code-explorer agent to understand current authentication flow Launch code-explorer to trace how authentication works ``` ## Key Source Files and Implementation Details The workflow automation is defined in the following locations within the `anthropics/claude-code` repository: - [`plugins/feature-dev/README.md`](https://github.com/anthropics/claude-code/blob/main/plugins/feature-dev/README.md) – Full description of the 7-phase workflow, philosophy, and best practices - [`plugins/feature-dev/commands/feature-dev.md`](https://github.com/anthropics/claude-code/blob/main/plugins/feature-dev/commands/feature-dev.md) – Command definition encoding phase actions for Claude Code - [`plugins/feature-dev/agents/code-explorer.md`](https://github.com/anthropics/claude-code/blob/main/plugins/feature-dev/agents/code-explorer.md) – Agent prompt and output format for exploration - [`plugins/feature-dev/agents/code-architect.md`](https://github.com/anthropics/claude-code/blob/main/plugins/feature-dev/agents/code-architect.md) – Agent prompt and output format for architecture design - [`plugins/feature-dev/agents/code-reviewer.md`](https://github.com/anthropics/claude-code/blob/main/plugins/feature-dev/agents/code-reviewer.md) – Agent prompt and output format for quality review - [`CLAUDE.md`](https://github.com/anthropics/claude-code/blob/main/CLAUDE.md) – Project style guide referenced by the code-reviewer agents during Phase 6 ## Summary - Claude Code's **7-phase feature development workflow** automates the path from informal request to production-ready code through deterministic stages. - **Three specialized agents**—`code-explorer`, `code-architect`, and `code-reviewer`—handle exploration, design, and validation in parallel where appropriate. - The workflow enforces **clarifying checkpoints** before design and **quality gates** before completion, reducing technical debt and rework. - All phase definitions and agent prompts are open-source in `plugins/feature-dev/`, making the automation transparent and customizable. - Developers interact via the `/feature-dev` command, retaining approval authority at architecture selection and quality review stages. ## Frequently Asked Questions ### How do I start the 7-phase feature development workflow in Claude Code? Invoke the `/feature-dev` command followed by your feature description. For example: `/feature-dev Implement dark mode toggle`. Claude Code will automatically proceed through Discovery, Exploration, Clarifying Questions, Architecture Design, Implementation, Quality Review, and Summary without additional commands. ### What is the difference between `code-explorer` and `code-architect` agents? The **`code-explorer`** agents investigate the existing codebase to locate relevant files and patterns, returning 5-10 key files to examine. The **`code-architect`** agents use those findings to generate implementation blueprints, proposing multiple architectural strategies (minimal change versus clean architecture) with explicit trade-off analysis for the developer to choose from. ### Can I skip phases or run individual agents manually? While the `/feature-dev` command orchestrates all seven phases sequentially, you can invoke individual agents manually using natural language commands like `Launch code-explorer to analyze the payment processing module`. However, skipping phases risks missing clarifying questions or architectural review, which the workflow is designed to enforce for quality assurance. ### Where are the agent prompts and workflow definitions stored? All definitions live in the `plugins/feature-dev/` directory. The README at [`plugins/feature-dev/README.md`](https://github.com/anthropics/claude-code/blob/main/plugins/feature-dev/README.md) documents the philosophy and phase details, while [`plugins/feature-dev/commands/feature-dev.md`](https://github.com/anthropics/claude-code/blob/main/plugins/feature-dev/commands/feature-dev.md) encodes the command logic. Individual agent prompts reside in [`plugins/feature-dev/agents/code-explorer.md`](https://github.com/anthropics/claude-code/blob/main/plugins/feature-dev/agents/code-explorer.md), [`code-architect.md`](https://github.com/anthropics/claude-code/blob/main/code-architect.md), and [`code-reviewer.md`](https://github.com/anthropics/claude-code/blob/main/code-reviewer.md).