Project infrastructure, agent orchestration, and architecture design for Claude Code.
Takes a vague idea through to implemented code via a pipeline of specialized, context-isolated skills — each handling one phase of the lifecycle.
/dk (type this — the router handles the rest)
│
├── dk-vision — conversational vision exploration with the user
├── dk-design — focused deep-dive sessions per area
├── dk-plan — automated domain-driven implementation planning
├── dk-review — conversational plan review and approval
└── dk-batch — automated serial stage execution
Key ideas:
- One entry point (
/dk) — the router reads file-based state and invokes the right skill - Conversational phases get fresh terminal sessions for clean context
- Automated phases delegate to sub-agents
- Crash recovery via progress markers — resume where you left off across sessions
- Spec system for maintaining living architecture documentation
git clone https://github.com/aahl-byte/ai-dev-kernel.git /tmp/ai-dev-kernel
/tmp/ai-dev-kernel/init.sh /path/to/your/project
rm -rf /tmp/ai-dev-kernelOr from within your project directory:
git clone https://github.com/aahl-byte/ai-dev-kernel.git /tmp/ai-dev-kernel
/tmp/ai-dev-kernel/init.sh .| Location | Contents |
|---|---|
.claude/skills/dk*/ |
Pipeline skills (router, vision, design, plan, review, batch, orchestrator, validate) |
.claude/skills/manage-specs/ |
Spec maintenance skill |
scripts/dk/launch.ts |
Fresh terminal launcher (macOS: Terminal.app, iTerm2, WezTerm) |
scripts/dk/validate.py |
status.yaml schema validator |
scripts/dk/ntfy.js |
Push notifications via ntfy.sh (set NTFY_TOPIC env var) |
scripts/spec-tree.py |
Spec tree viewer |
specs/ |
Pipeline specs + scaffolded ARCHITECTURE.md and INTENT.md |
- Claude Code CLI
- Node.js / tsx (for launch script)
- Python 3 (for spec-tree)
After installing, start Claude Code in your project and type:
/dk
The router will ask what you want to build and guide you through the pipeline.
| Phase | Skill | Mode | What happens |
|---|---|---|---|
| Vision | dk-vision | Conversational | Explores the big picture — goal, scope, areas |
| Design | dk-design | Conversational | Deep-dives each area that needs detail |
| Plan | dk-plan | Automated | Breaks design into staged implementation tickets |
| Review | dk-review | Conversational | You review, modify, and approve the plan |
| Execute | dk-batch | Automated | Implements stages one at a time |
Just type /dk again. The router reads state from tmp/initiatives/ and picks up where you left off — even across sessions.
When transitioning between conversational phases, the router opens a new terminal with a clean Claude session. This prevents context window pollution from long conversations.
Requires bun dk:launch (added to package.json during init).
The spec system maintains living documentation of your project's architecture. Two roots:
INTENT.md— what the project is and why (product philosophy)ARCHITECTURE.md— how it works (technical overview, parent of all technical specs)
Run python3 scripts/spec-tree.py (or bun specs:list if configured) to view the tree.
Use /manage-specs to audit and update specs after code changes.
After installing, you'll want to customize:
specs/INTENT.md— Your project's purpose and philosophyspecs/ARCHITECTURE.md— Your project's technical architecturedk-plantriage rules — Add project-specific deferral signals (what work should be flagged for manual handling)dk-validatedomains — Configure validation domains to match your project's architecturedk-orchestratorconventions — Add your project's type-check command, commit conventions, etc.
tmp/initiatives/{name}/
├── vision.md # Vision document
├── dive-tracker.yaml # State machine
├── design.md # Consolidated design
├── dives/{area}.md # Deep dive documents
├── classification.md # Domain classification
├── designs/{domain}.md # Domain designs (with Review + Verdict)
├── specs/{domain}.md # Condensed implementation specs
├── status.yaml # Execution plan
├── deferred.yaml # Deferred tickets
├── questions.md # Open questions
├── progress.plan # Planning progress marker
└── progress.review # Review progress marker
MIT