Orchestrating controlled agentic workflows.
A curated collection of agent-native workflow packages for software delivery, design, review, and automation. Each workflow is self-contained, installable into any repository, and engineered for controlled execution with validation gates, independent review, and bounded correction loops.
Controlled feature delivery through scoped requirements, cost-aware validation, tiered independent review, and focused delta corrections — from first line to ship-ready.
| Category | Software Engineering |
| Skills | 28 specialized skills |
| Roles | Router · Researcher · Architect · Builder · Integrator · Reviewer |
| State | Machine-checked via Python validator |
| Entrypoint | START_FEATURE_PROMPT.md |
Lifecycle:
IDLE → SCOPE → RESEARCH → [DESIGN] → BUILD → [INTEGRATE] → VALIDATE
→ REVIEW → [FIX → VALIDATE → REVIEW] → [HUMAN CHECKPOINT]
→ READY TO SHIP → SHIPPED
Key capabilities:
- Skill-routed execution with 28 engineering skills — debugging, TDD, code review, cost control, stakeholder requirements, spec-driven development, and more
- Risk-classified routing (LOW / MEDIUM / HIGH) with automatic review escalation
- Snapshot-based review tied to content — edits after review invalidate the verdict
- Bounded fix loops (default 2 cycles, then BLOCKED)
- Machine-checked readiness via
validate_workflow.py - Agent-agnostic — works with Claude, Codex, Gemini, or any capable engine
→ Documentation · Install
Design-specialized workflow with 15 skills and intelligent routing for building, redesigning, critiquing, and polishing frontend interfaces.
| Category | Design |
| Skills | 15 design skills + skill router |
| Routing | Automatic via skill-router |
| Entrypoint | AGENTS.md |
Lifecycle:
IDLE → ROUTE → RESEARCH → DESIGN → BUILD → VALIDATE → REVIEW → [FIX] → DONE
Available skills:
| Skill | Purpose |
|---|---|
web-design-engineer |
Full websites, dashboards, prototypes, major redesigns |
emil-design-eng |
Component polish, micro-interactions, UI craft |
animate |
Web motion, transitions, scroll/entrance animations |
landing-page-design |
Conversion landing pages, marketing sites |
build-awwwards-quality-sites |
Premium cinematic marketing sites |
video-to-superprompt |
Video reference → detailed implementation brief |
tastemaker |
Anti-generic visual direction, reference grounding |
better-layout |
Spacing, alignment, grouping, visual hierarchy |
interface-review |
UI diff and regression review |
critique-screen |
Rendered-screen visual critique |
perception-laws |
Gestalt, Fitts's Law, Hick's Law, Jakob's Law |
popular-web-designs |
54 real design systems as HTML/CSS reference |
uxpeak-ux-ui-design |
UX audit and redesign (product/conversion psychology) |
standalone-html-ui-artifacts |
Portable self-contained HTML deliverables |
iterative-ux-review |
Screenshot → analyze → fix loop |
→ Documentation · Install
agent-workflow/
├── workflows/
│ ├── feature-delivery/ 28-skill software engineering workflow
│ │ ├── README.md
│ │ ├── workflow.yaml
│ │ └── package/ installable files
│ └── frontend-ui-ux/ 15-skill design workflow
│ ├── README.md
│ ├── workflow.yaml
│ └── package/ installable files
├── shared/ shared resources (when extracted)
├── docs/ cross-cutting documentation
├── AGENTS.md repository rules
└── LICENSE MIT
Each workflow follows a standard layout:
workflows/<id>/
├── README.md workflow documentation
├── workflow.yaml metadata manifest
└── package/ files copied into the target repository
├── AGENTS.md policy for agents working under this workflow
├── .agent/ workflow internals, roles, skills, scripts
└── .active/ mutable runtime state (feature, review, decisions)
- Choose a workflow from the list above.
- Copy the contents of its
package/directory into your target repository root. - If the target repository already has an
AGENTS.md, merge the workflow rules into the existing file — do not overwrite project-specific policy.
Each workflow README has detailed installation instructions.
These principles are shared across all workflows:
- Bounded fix loops — maximum 2 broad fix/re-review cycles by default; after exhaustion, STOP and escalate to human.
- Independent review — Builder and Reviewer are separate contexts; the Builder may not approve its own work.
- Evidence over claims — validation requires fresh commands run after the last edit with exact output recorded; no fabricated results.
- Shipping is a human decision — agents do not commit, push, merge, or deploy unless explicitly authorized.
See Creating a Workflow and Workflow Conventions.