AI App Builder
Principle
Use AI intentionally only when it is central to the product. If AI is secondary, ask for approval before adding AI features. Keep the normal application complete: dashboard-first UI, real workflows, persistence, tests, auth after functionality, layered security, docs, and deployment.
AI Application Blueprint
- Define the core AI job: generation, extraction, classification, retrieval, recommendation, workflow automation, chat, agentic execution, or decision support.
- Define the user workflow around the AI job, including inputs, review, editing, saving, exporting, and history.
- Define required data: user content, prompts, model outputs, feedback, files, embeddings, tool results, cost/usage, and audit logs.
- Define model/provider requirements without hard-coding current model assumptions. Verify provider docs before implementation.
- Define safety boundaries: what the AI may do, what requires user approval, and what it must refuse or escalate.
- Define validation: schema checks, output checks, user confirmation, evals, and fallback behavior.
Build Order
- Build the main app workflow and dashboard shell.
- Build the AI happy path with controlled test data.
- Store prompts, inputs, outputs, status, usage, and feedback when persistence is required.
- Add prompt and output validation before saving important results.
- Add retries, timeouts, streaming, cancellation, and error states where relevant.
- Add auth, authorization, rate limits, usage quotas, and audit logs before stable release.
- Add evaluation cases for quality, reliability, safety, and regression.
- Document setup, environment variables, model/provider configuration, and operational limits.
Required AI Components
Include these when relevant:
- Prompt templates or prompt builders
- Structured input and output schemas
- Model/provider adapter boundary
- Tool/function contracts
- RAG ingestion, retrieval, and citation path
- Conversation or job history
- Human review and approval flow
- Feedback capture
- Usage and cost tracking
- Safety and abuse controls
- Evaluation fixtures and expected behavior
Ask Before Implementing
Ask before using external AI APIs, paid models, persistent user data, scraping, file ingestion, agentic actions, email/message sending, payments, or anything that changes external systems.
Completion Check
AI app work is not done until:
- The core AI workflow works end to end.
- Outputs are validated before critical use.
- Failures are visible and recoverable.
- Required data is stored correctly.
- Cost, rate, and safety boundaries are defined.
- Verification and validation both pass.