ai-prd-workflow

Orchestrates a sequential chain of specialized prompts that progressively refine vague product ideas into structured RFCs. Each stage (clarification → analysis → specification → implementation) feeds outputs as context into the next stage, creating a dependency graph where later prompts leverage earlier structured outputs. The pipeline is agnostic to the underlying LLM, accepting any AI assistant via standard text interfaces (Claude, ChatGPT, Cursor, etc.).

The first stage of the pipeline uses targeted prompts to extract and clarify implicit assumptions, ambiguities, and scope boundaries from a vague product idea. It systematically questions the idea across dimensions (user personas, success metrics, constraints, dependencies) and produces a structured clarification document that serves as the foundation for all downstream stages. This stage acts as a requirements elicitation engine, converting narrative descriptions into enumerated, unambiguous statements.

Takes the clarified requirements and performs a structured technical analysis to identify architectural patterns, technology choices, potential bottlenecks, and implementation risks. This stage synthesizes the clarification output with technical knowledge to produce a feasibility assessment and high-level architecture recommendation. It operates as a technical advisor layer, evaluating trade-offs between different implementation approaches and flagging risks early.

Synthesizes outputs from clarification and technical analysis stages to generate a complete, structured RFC document with detailed specifications, acceptance criteria, and implementation guidelines. This stage uses a template-driven approach where the prompt includes a specification schema (sections for overview, requirements, architecture, acceptance criteria, timeline, dependencies) and fills each section with content derived from earlier stages. The output is formatted for direct consumption by developers and code generation tools.

Breaks down the RFC into granular, sequenced implementation tasks with estimated effort, dependencies, and success criteria. This stage takes the detailed specification and produces a task list that developers can immediately begin working from, including task ordering based on dependencies, effort estimates, and clear acceptance criteria for each task. It operates as a project planning layer, converting specification into actionable work items.

Provides a shell-based execution framework that chains prompts across different LLM providers (Claude, ChatGPT, Cursor, Ollama) without requiring SDK-specific code. The pipeline uses standard input/output redirection and API calls to invoke different LLMs, storing intermediate outputs as files that feed into subsequent stages. This architecture enables users to mix and match LLM providers (e.g., use Claude for clarification, GPT-4 for analysis, Cursor for code generation) without rewriting the pipeline.

Implements a prompt chaining pattern where each stage's output is automatically included as context in the next stage's prompt, creating a dependency graph of prompts. The pipeline uses file-based context passing where outputs from stage N become inputs to stage N+1, enabling later stages to reference and build upon earlier structured outputs. This pattern reduces hallucination and improves coherence by ensuring each stage operates on concrete, structured context rather than abstract requirements.

Provides a structured checkpoint system that formalizes 'vibe coding' workflows (rapid prototyping with AI assistants) by injecting specification and planning stages between ideation and implementation. The pipeline acts as a formalization layer that captures the implicit decisions made during vibe coding and converts them into explicit, documented specifications. This enables teams to maintain the speed of vibe coding while adding rigor and traceability.