Human-first AI-driven development workflow for Claude Code.
A Claude Code plugin that transforms AI coding from "execute my spec" into collaborative building — where you drive decisions and Claude augments your capabilities.
Traditional AI coding workflows treat you as a spec writer:
- You write detailed requirements upfront
- AI executes without understanding context
- Changes feel like starting over
- No learning between sessions
Codie flips the dynamic. You're the visionary, Claude is the builder:
- Collaborative understanding - Start with problems, not specs
- Ongoing dialogue - Syncs feel like check-ins, not approval gates
- Learning-focused - Capture insights, improve continuously
- Milestone-based - Ship value incrementally
# Add the marketplace
/plugin marketplace add AI-Native-Systems/codie-cc-plugins
# Install the plugin
/plugin install codie@ans
# Start a new project
/codie:start| Command | Description |
|---|---|
/codie:start |
Begin project through collaborative understanding |
/codie:discuss |
Explore possibilities and shape understanding |
/codie:plan |
Shape an approach with intentions and flexibility |
/codie:execute |
Build together with ongoing dialogue |
/codie:verify |
Reflect on what was built and chart improvement path |
/codie:debug |
Systematic investigation |
/codie:quick |
Fast path for simple tasks |
/codie:map |
Understand existing codebase |
/codie:help |
Show all commands |
/codie:start → Understand the problem, shape a vision
/codie:discuss → Explore possibilities for a milestone
/codie:plan → Shape an approach together
/codie:execute → Build with ongoing dialogue
/codie:verify → Reflect and identify improvements
Loop: discuss → plan → execute → verify (repeat per milestone)
Quick tasks: /codie:quick for ad-hoc work without full planning.
Instead of writing specs, have a conversation:
- What problem are you solving?
- Who is it for?
- What does success look like?
Creates: vision.md, journey.md, memory.md, learnings.md
Optional: requirements.md (formal tracking), research/ (domain research)
For each milestone, explore options:
- What approaches are possible?
- What trade-offs exist?
- What have we learned?
Creates: exploration.md for the milestone
Turn exploration into action:
- Work breakdown with intentions (not rigid specs)
- Sync points (not approval gates)
- Flexibility for discoveries
Creates: approach.md files
Collaborative building with ongoing dialogue:
- Discoveries are celebrated, not flagged as deviations
- Memory updated as context evolves
- Stories capture what was built and why
Creates: story.md files, updates memory.md
Collaborative reflection (not pass/fail judgment):
- What worked well?
- What could be better?
- What opportunities emerged?
Creates: reflection.md, updates learnings.md
.planning/
├── config.json # Workflow preferences
├── vision.md # Shared understanding
├── journey.md # Value milestones
├── memory.md # Project memory & state
├── learnings.md # Accumulated insights
├── understanding/ # Codebase analysis (from /codie:map)
└── milestones/ # Milestone work
└── 01-{name}/
├── exploration.md
├── 01-approach.md
├── 01-story.md
└── reflection.md
- Human-first: You drive decisions, Claude augments your capabilities
- Design-driven: Start with problems and people, not feature lists
- Learning-focused: Capture insights, improve continuously
- Collaborative: Ongoing dialogue, not checkpoint rituals
Working with Codie feels:
- Empowering — you make the calls, AI handles the execution
- Curious — discoveries are celebrated, not flagged as deviations
- Progressive — each milestone builds on learnings
- Conversational — syncs feel like check-ins, not approval gates
Codie is most valuable for the intermediate user with a non-trivial project — someone who:
- Knows Claude Code basics
- Is building something that spans multiple sessions
- Has experienced context drift frustration
- Wants structure but not rigidity
- No external dependencies - All functionality is self-contained
- No data collection - Your context stays local
- No network calls - Except for Claude API (handled by Claude Code)
- Open source - Full transparency
- Auto-commit code (user controls git)
- Send data to external services
- Access files outside project directory
- Run commands without user visibility
- Bypass Claude Code's permission system
- Claude Code 1.0 or later
- No external dependencies
Initial design was based on open source contributions from Lex Christopherson.
Contributions welcome! See CONTRIBUTING.md for guidelines.
MIT License - see LICENSE for details.
Built for the ans marketplace
Built by AI Native Systems™
