feature-factory-core

feature-factory

OtherClaude Codeby kororos

Summary

Stack-agnostic multi-agent feature-delivery chain: codebase-researcher -> story-writer -> spec-writer -> backend-builder -> frontend-builder -> test-verifier -> e2e-author -> implementation-validator, with human approval gates and a validator loop-back. Includes the feature-factory skill and a secret-scan pre-commit hook.

Install to Claude Code

/plugin install feature-factory-core@feature-factory

Run in Claude Code. Add the marketplace first with /plugin marketplace add kororos/cc-feature-factory if you haven't already.

README.md

feature-factory — Claude Code plugin marketplace

A small marketplace with three plugins that package the multi-agent feature-delivery workflow so it can be reused across any project and shared with teammates.

  • feature-factory-core — stack-agnostic. The eight specialist agents, the

feature-factory orchestration skill, and a secret-scan pre-commit hook. Its end-to-end step is UI-runtime-neutral: a browser for web apps, a simulator/emulator for mobile apps.

  • nextjs-prisma-overlay — optional. A stack-conventions skill encoding the

Next.js (App Router) + TypeScript + Prisma/PostgreSQL + BullMQ/Redis + Resend + Docker/Playwright conventions. Enable it only on matching repos; it sharpens the core agents with that stack's build/test/migration/deploy rules.

  • react-native-overlay — optional. A stack-conventions skill encoding the

React Native (Expo managed/dev-client or bare RN) + TypeScript conventions: navigation, client state, the data layer, native config/permissions, on-simulator verification, and the e2e harness. It detects the RN flavor and the e2e harness (Detox/Maestro), asking only when the repo is ambiguous. Enable it only on matching repos.

The core agents discover each project's stack (from the manifest — package.json,

pyproject.toml, go.mod, … — and the project's CLAUDE.md) and, if the

stack-conventions skill is installed, consult it. So the core works anywhere; the overlay makes it stack-aware where it applies.

What's inside

.claude-plugin/marketplace.json          # lists both plugins
plugins/
  feature-factory-core/
    .claude-plugin/plugin.json
    agents/                              # codebase-researcher, story-writer, spec-writer,
                                         # backend-builder, frontend-builder, test-verifier,
                                         # e2e-author, implementation-validator, feature-orchestrator
    skills/feature-factory/SKILL.md      # the 8-step chain + 3 approval gates + validator loop-back
    hooks/hooks.json + pre-commit.sh     # blocks commits containing .env/.key/.pem/secrets/creds
  nextjs-prisma-overlay/
    .claude-plugin/plugin.json
    skills/stack-conventions/SKILL.md
  react-native-overlay/
    .claude-plugin/plugin.json
    skills/stack-conventions/SKILL.md   # Expo/bare RN; detects flavor + e2e harness

The chain

codebase-researcher → story-writer → (approval) → spec-writer → (approval) → backend-builder → frontend-builder → test-verifier → e2e-author → implementation-validator → (approval), looping back to the right builder if the validator finds critical gaps. Run it with the feature-factory skill (or the

feature-orchestrator agent), or invoke any agent on its own.

Install (in any project)

# From a local clone of this repo:
/plugin marketplace add /path/to/feature-factory
# …or from a git remote once you push it:
/plugin marketplace add <git-url>

/plugin install feature-factory-core@feature-factory
# On Next.js + Prisma repos, also:
/plugin install nextjs-prisma-overlay@feature-factory
# On React Native (Expo or bare) repos, instead:
/plugin install react-native-overlay@feature-factory

Enable at most one stack overlay per repo — each ships a skill named

stack-conventions, and the core agents consult whichever one is installed.

Enable per project by committing enabledPlugins to the project's .claude/settings.json, or globally in ~/.claude/settings.json.

Deliberately NOT included

  • Agent memory (.claude/agent-memory/*) — these are per-project learnings

(database/container names, this-repo gotchas). They must NOT travel between projects; each repo accrues its own. The memory: project setting on each agent keeps memory scoped to the project it runs in.

  • MCP servers / connectors (telegram, Context7, Gmail, etc.) and the global

voice-transcription CLAUDE.md — these are already user-scoped (in

~/.claude/settings.json enabledPlugins and ~/.claude.json), so they apply to every project automatically. Nothing to package here. Environment-specific MCP (e.g. a docker-socket proxy for one dev box) is intentionally left out.

Provenance

Extracted and genericized from the notesapp project's .claude/ setup. The core agents/skill had all notesapp-specific names (container/DB names, the tsvector gotcha, exact compose paths) stripped; those repo-specific details belong in each project's own CLAUDE.md.

Related plugins

Browse all →