AI Phase Transitions & Spec Generation

[tkowalczyk]

Parent PRD

#1

What to build

Extend the AI system prompts so Claude auto-transitions through four phases: discovery → requirements (PRD) → implementation plan → task breakdown. Each transition is clearly communicated to the client in plain, non-technical language.

System prompts: Encode the full interview methodology. The AI should know when discovery is sufficiently complete to move on, synthesize findings into a structured requirements doc, break that into a phased plan, and finally produce a task list.

Output parsing: When the final phase completes, parse the AI's structured output into a markdown spec and individual tasks. Store the spec in the Spec table and tasks in the Task table. Update project status to "review."

Context management: For long conversations spanning multiple days, implement summarization when approaching context limits — summarize earlier phases and inject the summary so the AI maintains coherence.

Prompt storage: System prompts should be stored in config/DB, not hardcoded in source, so the methodology can be iterated without code changes.

Acceptance criteria

• AI transitions from discovery → requirements → plan → tasks automatically
• Each phase transition is clearly signaled to the client in plain language
• Final output contains a structured markdown spec and numbered task list
• Spec is parsed and stored in the Spec table
• Tasks are parsed and stored individually in the Task table
• Project status updates to "review" on completion
• Long conversations (50+ messages) maintain coherence via summarization
• System prompts are configurable without code changes

Blocked by

• Blocked by Client Chat Interface & AI Integration #3

User stories addressed

• User story 4
• User story 19
• User story 20
• User story 21
• User story 22
• User story 23