Exponential development workflow — toolkit composed by skills, agents, and architecture rules for Claude Code, Cursor, Copilot and more.
We're in a new paradigm. AI agents can write code — but they need structure to do it well. Without clear patterns, quality gates, and workflow guidance, you end up reviewing, correcting, and redoing the work you delegated.
The developer's job is to define the system, focus on the problem to solve, and own the experience. The AI's job is to execute. This creates a perfect division of responsibilities — each actor focused on where they add the most value — and that's where the exponential development starts.
devtronic is that structure. A toolkit composed by skills, agents, quality gates, and self-improving rules — tailored to your stack. Refined daily across real projects, personal and enterprise.
Not a silver bullet. A starting point built from experience, open to iteration.
Works with Claude Code, Cursor, Google Antigravity, GitHub Copilot, and OpenCode.
| Document | Description |
|---|---|
| Tutorials | Step-by-step guides for common use cases |
| Skills Reference | Detailed documentation of all 21 core + 12 design phase + 9 addon skills |
| Agents Reference | Detailed documentation of all 15 core agents + 4 auto-devtronic addon agents |
| Plugin Mode | Claude Code plugin architecture, hooks, and migration |
| CLI Reference | Full command documentation |
| Existing Projects | Integration with existing configurations |
| Customization | How to customize rules, skills, agents |
| Philosophy | Why this workflow works |
| Worktrees Guide | Parallel Claude sessions |
| Design Phase Guide | Design phase skills, agents, and workflow |
| Multi-Account Setup | Multiple Claude Code accounts |
| auto-devtronic Addon | Autonomous engineering loop addon |
# Run in your project directory
npx devtronic init
# Or specify a target directory
npx devtronic init /path/to/your/projectThe CLI will:
- Analyze your project (framework, architecture, stack)
- Ask to confirm or adjust the detected configuration
- Select which IDEs to configure
- Offer optional addon packs via multiselect (Claude Code)
- Generate personalized rules, skills, agents, and hooks
- Track installation for future updates
/setup # For existing projects
/scaffold # For new projects from scratch
Same functionality through conversation. Use /scaffold to create new projects with guided architecture selection (frontend, spa-ddd, backend, monorepo).
In Claude Code, run:
/plugin marketplace add r-bart/devtronic-plugin
/plugin install devtronic@devtronic
This installs skills, agents, and hooks directly from the marketplace — no CLI needed. Use the CLI (npx devtronic init) for full setup including rules, AGENTS.md, and stack detection.
npx devtronic update # Apply updates
npx devtronic update --check # Check for updatesIf the package isn't published to npm yet, you can use it locally:
# 1. Clone and build
git clone https://github.com/r-bart/devtronic.git
cd devtronic/packages/cli
npm install && npm run build
npm link
# 2. Use in any project
cd /path/to/your/project
npx devtronic init| IDE | Config Location | Status |
|---|---|---|
| Claude Code | GitHub marketplace + .claude/rules/ |
Full support (plugin mode) |
| Cursor | .cursor/rules/*.mdc |
Full support |
| Google Antigravity | .agents/rules/*.md |
Full support |
| GitHub Copilot | .github/copilot-instructions.md |
Full support (enable in GitHub settings) |
| OpenCode | .opencode/rules/*.md |
Full support |
| Zed | Uses AGENTS.md directly |
Native |
┌─────────────────────────────────────────────────────────────────┐
│ AI ARCHITECTURE LAYERS │
├─────────────────────────────────────────────────────────────────┤
│ │
│ AGENTS.md Universal context for all AI agents │
│ │ │
│ ├── Skills Reusable workflows (/spec, etc.) │
│ │ 21 core + 12 design phase + 9 addon │
│ │ │
│ ├── Agents Specialized subagents (quality, review) │
│ │ 15 core + 4 addon agents │
│ │ │
│ ├── Rules Quality standards (IDE-specific format) │
│ │ │
│ └── Hooks Automated workflow (lint, checkpoint, etc.) │
│ 5 hooks included (Claude Code) │
│ │
│ thoughts/ Persistent documents (specs, plans, etc.) │
│ │
└─────────────────────────────────────────────────────────────────┘
/brief [topic] → Orientation + pre-flight checks
│
[ /design phase ] → (optional) UX discovery → wireframes → design system
│ See Design Phase Guide ↗
/spec [idea] → Define WHAT to build (PRD)
│
/create-plan [feature] → Design implementation phases
│
/generate-tests → Encode acceptance criteria as failing tests
│
/execute-plan → Implement in parallel phases
│
[ /design-review ] → (optional) QA vs wireframes + design system
│
/summary → Document what changed and why
│
/post-review → Final quality check
Human review at earlier stages has higher leverage. See Philosophy for details.
For UI-heavy features, run the design phase before /spec. See Design Phase Guide for the full UX workflow.
Tired of driving every step? The Convergence Loop (
/loop) runs the middle of this pipeline autonomously — you sign only the DoD and the ship.
The skills above work individually — run /spec, /create-plan, /execute-plan by hand, one turn at a time. The convergence loop is the upgrade: it keeps a human only at the two ends — you sign the DoD (the tests) up front and the ship at the back — and lets the machine converge everything in between under gates that never tire (the barbell).
/spec → /generate-tests → you sign the DoD (the tests)
/loop <feature> → machine converges: Tier ① gates every iteration,
Tier ② adversarial review at barriers, bounded by budget
→ STOPS at the ship for your QA + signature
Driven by a per-repo loop.manifest.yaml (seeded by init, fully commented). Preview it with devtronic loop --dry-run — it explains the plan without executing anything. Inert by default: with no manifest, the ambient hooks behave exactly as before.
The loop of loops: point it at your /backlog (items with a - Spec: + - DoD:), walk away, and come back to a queue of converged features ready to sign — each in its own git worktree, bounded by a width cap + token budget, fail-soft (a stuck item is quarantined; the run continues).
devtronic loop --backlog --dry-run # preview the eligible queue + caps
/loop --backlog # converge each item, park for your sign
devtronic loop --backlog --status # the parked sign-queue
devtronic loop --backlog --sign BACK-001 # QA the worktree, then shipRequires the
devtronicCLI on your PATH — the/loopskill and theStophook shell out todevtronic loop …. Install it globally:npm i -g devtronic. Without it the individual skills still work; only the loop does not.
Not to be confused with the auto-devtronic addon (
/devtronic, below) — a separate GitHub-issue-to-PR pipeline.
See the CLI reference and Skills reference for the full contract.
| Category | Skills |
|---|---|
| Orientation & Research | /brief, /research, /opensrc |
| Planning | /spec, /create-plan |
| Development | /scaffold, /setup, /investigate, /worktree |
| Execution | /quick, /execute-plan, /loop |
| Quality & Review | /audit, /post-review, /generate-tests |
| Session & Meta | /checkpoint, /summary, /backlog, /learn, /create-skill, /devtronic-help |
| Design Phase | /design, /design-research, /design-define, /design-ia, /design-wireframe, /design-system, /design-system-define, /design-system-audit, /design-system-sync, /design-audit, /design-review, /design-spec |
| Orchestration (addon) | /briefing, /recap, /handoff |
| Design Best Practices (addon) | /design-init, /design-critique, /design-refine, /design-tokens, /design-harden |
| Auto-devtronic (addon) | /devtronic — Autonomous engineering loop |
See Skills Reference for detailed documentation of each skill.
The auto-devtronic addon runs the full pipeline autonomously — spec → tests → plan → implement → PR — and self-corrects via failing tests.
npx devtronic addon add auto-devtronic
/devtronic https://github.com/org/repo/issues/42 # HITL mode (default, human gates)
/devtronic <issue> --afk # AFK mode (fully autonomous)
/devtronic <issue> --afk --validate # Validate task first, then runTwo modes:
| Mode | Flag | Human gates |
|---|---|---|
| HITL | --hitl (default) |
Pauses at spec, tests, each retry, and PR |
| AFK | --afk |
Fully autonomous, no interruptions |
Set the default mode project-wide:
npx devtronic mode afk # set AFK as default
npx devtronic mode hitl # set HITL as default
npx devtronic mode show # show current modeSee auto-devtronic Addon for full documentation.
| Agent | Model | Purpose |
|---|---|---|
| error-investigator | haiku | Quick automatic error diagnosis |
| code-reviewer | sonnet | Thorough PR review |
| architecture-checker | sonnet | Validate Clean Architecture compliance |
| quality-runner | haiku | Run tests, typecheck, and lint |
| commit-changes | haiku | Atomic conventional commits |
| test-generator | sonnet | Generate unit tests following project patterns |
| dependency-checker | haiku | Audit dependencies for vulnerabilities and issues |
| doc-sync | haiku | Verify docs match the actual codebase |
| ux-researcher | sonnet | Synthesize research, personas, user journeys |
| ia-architect | sonnet | Navigation structure, user flows, sitemaps |
| design-critic | sonnet | Nielsen's 10 heuristics evaluation |
| a11y-auditor | haiku | WCAG 2.1 AA accessibility compliance |
| design-token-extractor | haiku | Extract and normalize design tokens |
| design-system-guardian | haiku | Detect design system drift (read-only) |
| visual-qa | sonnet | Compare implementation vs design specs |
See Agents Reference for detailed documentation of each agent.
After setup, your project will have:
your-project/
├── AGENTS.md # Universal AI context (customize this!)
├── CLAUDE.md # Claude Code project rules
│
├── .claude/ # Claude Code configuration
│ ├── rules/ # Auto-applied architecture rules
│ └── settings.json # Marketplace registration
│
├── .cursor/rules/ # Cursor rules (if selected)
├── .agent/rules/ # Antigravity rules (if selected)
│
├── .ai-template/ # Installation manifest
│ └── manifest.json
│
└── thoughts/ # AI working documents
├── specs/ # PRDs from /spec
├── research/ # Research from /research
├── plans/ # Plans from /create-plan
├── summaries/ # Change summaries from /summary
├── checkpoints/ # Auto-checkpoints before compaction
├── notes/ # Project notes
├── debug/ # Debug analysis
├── audit/ # Audit reports from /audit
├── design/ # Design artifacts
└── archive/ # Archived items
Skills, agents, and hooks are loaded from the GitHub marketplace plugin — no plugin files in your project.
See Plugin Mode for the full architecture and how hooks work.
The CLI is designed for projects that already have AI configuration. When conflicts are detected, you choose:
- Merge - Add new sections, keep existing customizations
- Keep - Skip files that already exist
- Replace - Overwrite with template files
See Existing Projects Guide for detailed integration instructions.
Every task should leave the codebase working:
- Tests green, lint clean, types valid after EVERY task
- If a task is too large, break it into smaller subtasks
After any correction:
"Update CLAUDE.md so you don't make that mistake again."
Claude is good at writing rules for itself.
Run 3-5 Claude sessions simultaneously using git worktrees. See Worktrees Guide.
| Command | Description |
|---|---|
init [path] |
Initialize in a project |
update |
Update to latest template |
add <ide> |
Add another IDE |
addon add <name> |
Add an optional skill pack |
addon remove <name> |
Remove an optional skill pack |
addon list |
List available and installed addons |
addon sync |
Regenerate addon files for current agent config |
regenerate |
Regenerate files |
status |
Show installation status |
diff |
Show differences with template |
info |
Version, config, and install summary |
list [skills|agents] |
List installed skills and agents |
config |
View or manage project configuration |
doctor [--fix] |
Run health diagnostics |
loop [--dry-run] |
Convergence loop (inner) — validate/preview loop.manifest.yaml |
loop --backlog |
Loop of loops (outer) — drive the /backlog queue unattended |
uninstall |
Remove devtronic from your project |
See CLI Reference for full documentation.
devtronic is designed to be extended, not forked.
Before opening a PR, consider whether your contribution can be packaged as:
- A custom skill (
.claude/skills/) — for new workflows - A custom agent (
.claude/agents/) — for specialized automation - A plugin — for distributing skills + agents + hooks as a package
PRs are welcome for:
- Bug fixes and correctness improvements
- Structural or architectural enhancements
- Documentation improvements
- CI/tooling improvements
For feature ideas that extend functionality, we strongly encourage publishing them as standalone skills or plugins that others can install. This keeps the core lean and the ecosystem modular.
See Customization for how to create skills, agents, and plugins.
MIT - Use and adapt freely for your projects.