feature-dev for Claude Code: review, how it works and whether it is worth it (2026)

What you get in one minute#

feature-dev is Anthropic's free plugin that turns one command, /feature-dev, into a seven-step routine (explore the code, ask questions, design, build, review), and it is worth it for multi-file features but not for small fixes.

Logo: the official Claude Code logo from Anthropic's press kit (anthropic.com/press-kit)

The marketplace page shows 256,017 installs (October 2026). It is small: one prompt file and three subagent files, no scripts, no hooks.

What it does#

You type /feature-dev Add rate limiting to the API and Claude stops being a code writer first: it explores, asks, proposes designs and only then codes. It waits for you at three points.

Added by the plugin:

• 1 slash command: /feature-dev (optional description of the feature).
• 3 subagents: code-explorer (traces how existing code works), code-architect (writes a build blueprint), code-reviewer (finds bugs, reports only confidence 80 or higher). You can also call each by hand ("launch code-reviewer on my changes").
• 0 skills, 0 hooks, 0 MCP servers.

How it works in the background#

One long prompt is the whole engine: Claude Code loads it as instructions, and it tells Claude when to launch subagents and when to stop and ask you. The picture follows one feature.

• Your three stops: answer the questions (step 6), pick an approach (step 8), decide what to fix (step 11).
• Parallel agents: the prompt asks for 2-3 explorers, 2-3 architects and 3 reviewers, so 7 to 9 subagent runs.
• Writing code happens in your main session, not in a subagent.

It is all written instructions; nothing forces Claude to follow a phase.

Under the hood: dissection#

feature-dev is 100% Markdown: no code at all runs by itself, so there is nothing to execute except what Claude decides to do with the prompt.

The file tree#

feature-dev/
├── .claude-plugin/
│   └── plugin.json          name, description, author (no version)
├── commands/
│   └── feature-dev.md       the 7-phase prompt, 125 lines
├── agents/
│   ├── code-explorer.md     yellow, model: sonnet
│   ├── code-architect.md    green, model: sonnet
│   └── code-reviewer.md     red, model: sonnet
├── LICENSE                  Apache 2.0
└── README.md                documentation only, 412 lines

What each file is#

Key excerpts#

commands/feature-dev.md, Phase 3:

**CRITICAL**: This is one of the most important phases. DO NOT SKIP.

agents/code-reviewer.md, the header and the filter:

tools: Glob, Grep, LS, Read, NotebookRead, WebFetch, TodoWrite, WebSearch, KillShell, BashOutput
model: sonnet
...
**Only report issues with confidence ≥ 80.**

Disk and network#

• Writes: nothing itself. The code you approve is written by Claude with your normal permissions.
• Network: no calls from the plugin; the subagents list WebFetch and WebSearch, which Claude may use.
• Needs: a git repository (the reviewer looks at git diff by default) and an existing codebase.

Chart: word counts of the files in plugins/feature-dev at commit 41f619d of anthropics/claude-plugins-official, read 27 October 2026. Only the grey README is never loaded.

What users say#

19 user reports read: 3 praise, 4 mixed, 12 complaints, but the split is skewed. GitHub issues attract complaints. Reddit could not be read, so it is not counted; one blog (Medium) returned an error and is not counted either.

Chart: our count of the 19 reports, October 2026. Not a satisfaction rate.

Chart: themes of the 12 complaints, labelled by DIY Lab Hub, October 2026.

Pros: why it is useful#

It forces the thinking that vague prompts skip, and it costs you nothing but usage.

• Questions before design: Phase 3 is marked "DO NOT SKIP", so Claude cannot guess your requirements.
• Three designs to compare (minimal, clean, pragmatic) with a recommendation, before any code.
• Review with a threshold: the reviewer reports only confidence 80 or higher, so fewer nitpicks.
• Nothing to trust blindly: no scripts or hooks; you can read all 256 lines of prompt in minutes.
• Reusable parts: the three agents work on their own, useful even if you never run the full flow.

Cons: why it is not useful#

It is a heavy routine with a pinned cheaper model, and it cannot tell that your task is small.

• Cost and speed: up to 9 subagent runs per feature, and complaints about slowness and context.
• Sonnet is hard-coded (model: sonnet). The docs say an omitted model follows your session; here it does not.
• Stale agent definitions: the tool list has no Bash and a reporter says four names no longer resolve; the plugin is unchanged since 2025.
• Can override your limits: the "launch 3 agents" text beat a user's "at most 1 agent" rule, and the reviewer names CLAUDE.md, not AGENTS.md.
• Overlap with plan mode and the code-review plugin; no version number, so update notices can repeat.

Should you consider it? When it is useful#

Consider it for features that touch several files or start vague; skip it for fixes, and try it on one real feature first.

Install and check it#

You install it from Anthropic's official marketplace with one command, inside Claude Code or from your shell.

Linux#

Inside Claude Code:

/plugin install feature-dev@claude-plugins-official

Or from your shell:

claude plugin install feature-dev@claude-plugins-official

Windows (PowerShell)#

No Bash or Git hook is needed (the plugin has no scripts), but the reviewer expects a git repository:

claude plugin install feature-dev@claude-plugins-official

Mac (Terminal)#

Same command as Linux:

claude plugin install feature-dev@claude-plugins-official

Check it worked#

claude plugin list
claude plugin details feature-dev

Then open a git project and type /feature-dev with a small idea; Claude should create a to-do list and ask what you want before it reads any code.

Things to keep in mind#

• Security. A plugin can change how Claude behaves. This one has no hooks, scripts or MCP servers, and the subagents lack Write; the code is written by your main session under your usual permission prompts.
• Usage. Parallel Sonnet subagents add up on big repositories; start with one real feature.
• Updates. With no version number, Claude Code compares commit IDs, so it may say "updated" without changes (issue 3571).
• Conflicts. Overlaps with superpowers and code-review; install one workflow plugin at a time.
• Organisations can limit marketplaces; if install is refused, ask yours.

Troubleshooting#

FAQ#

Is feature-dev made by Anthropic? Yes. It lives in Anthropic's own claude-plugins-official repository, author "Anthropic".

Can I use only the agents? Yes. Ask Claude to "launch code-explorer to trace how login works" or "launch code-reviewer on my changes".

Can I make the agents use Opus? Not without editing: the model is set in each agent file. Copy a file to your own agents folder and change model:.

How is it different from plan mode? Plan mode makes one plan; this runs separate exploration, design and review steps and stops for your answers between them.

Sources#

• Plugin files: command, three agents, README, plugin.json, read at commit 41f619d (27 October 2026); marketplace.json entry.
• feature-dev on the Claude marketplace: install count and listed components (October 2026).
• Official Claude Code docs: Create custom subagents, Plugin commands reference, Measure plugin cost, Plugin security and trust (October 2026).
• User reports: issues 3571, 163, 383, 4235, 256, 294, 6307 and 75066, 64395, 30845, 30843, 24932, 13968, 15849 (December 2025 to September 2026); Hacker News 46929320, 47395989, 46975980 (February to March 2026); Tom Ashworth (December 2025); Leanware (updated September 2026).

Not run by DIY Lab Hub: we read the files and docs but did not install the plugin on a test machine.

Checked against the official docs on 2026-10-27