How It Works

When you describe what you want to build, SKYCOT's intent extractor parses your description to identify core features, user roles, data entities, and key workflows. This stage runs on Claude Opus to ensure nuanced understanding of complex requirements.

The extractor produces a structured intent document that captures what your app does, who uses it, and what data it manages — all before any code is generated.

The intent document is matched against pre-configured app architectures using confidence-based routing. Three outcomes are possible:

• Strong match (high confidence) — A single archetype handles the build. Its pre-configured schemas, routes, and components are used as the foundation, cutting required questions by 70-80%.
• Partial match (medium confidence) — Multiple archetypes are detected. SKYCOT composes a hybrid template by blending features from each matching archetype.
• No match (low confidence) — The Custom fallback activates. SKYCOT asks more detailed questions and builds from first principles, without archetype assumptions.

Once the archetype is selected, smart follow-up questions fill in the remaining 20-30% of decisions that the template cannot infer — things like colour preferences, specific business rules, custom fields, or third-party integrations.

Questions are archetype-aware: a Marketplace archetype asks about commission rates and seller verification, while a Booking archetype asks about cancellation policies and timezone handling. This keeps the conversation short and relevant.

The compiler breaks the build into a directed acyclic graph (DAG) of sessions. Each session is responsible for a specific layer of the application — database schema, API routes, UI components, authentication, payments, and so on.

Sessions that have no dependencies on each other run in parallel, significantly reducing total build time. The orchestrator manages execution order, ensuring that dependent sessions (like API routes that depend on the database schema) wait for their prerequisites.

Each session produces a structured JSON checkpoint containing the files created, exports, endpoints, and migrations. These checkpoints are 95% smaller than full conversation state and enable resume-from-checkpoint on failure.

SKYCOT follows the AgentCoder pattern: for each session, an independent test agent writes tests first, then a code agent implements until all tests pass. This inverted workflow catches errors during generation rather than after, achieving production-quality output compared to roughly 60% for traditional generate-then-test approaches.

The test agent operates independently from the code agent to prevent confirmation bias — tests define the expected behaviour objectively, and the code agent is measured against that specification.

After code generation, a review agent checks each session for code quality, consistency with the overall architecture, and adherence to the project's coding standards. A security scanner then runs 10 automated checks covering hardcoded secrets, SQL injection, XSS vulnerabilities, missing auth guards, and more.

Builds that fail verification enter an automatic recovery loop: the orchestrator retries the failed session with an escalated model (Opus) and additional context about the failure. If recovery also fails, the build is flagged for manual review.

Once all sessions pass verification, your app is ready for live preview in the browser via WebContainers. When you are satisfied, one-click deployment pushes the application to Vercel (frontend and serverless functions) and provisions a Supabase project (PostgreSQL database, authentication, and storage).

The deployed application is fully independent — you receive the complete source code and can continue development outside of SKYCOT at any time.