Prompt System and AI Collaboration

1. Prompt System Overview

This chapter explains the prompt-driven collaboration mechanism used during the R&D phase of the A2E project. The system is used to organize requirement clarification, technical decomposition, implementation progress, and delivery review. It is not part of the A2E runtime engine itself.

The Prompt system consists of two types of documents:

• Guide: defines stage goals, boundaries, inputs, outputs, and completion criteria.
• Template: defines the structure of stage documents and constrains record format and deliverables.

The current system covers three starting points and five main stages:

• Starting points: Story / Feature / Issue
• Main flow: Draft -> Plan -> Code -> Iteration -> Review

2. Document-driven Development Flow

The goal of this flow is to gradually turn an "idea" into a "verifiable delivery result", while leaving a traceable record at each step.

2.1 Story

Story / Feature / Issue serves as the entry to the flow and defines "what problem this round is solving".

• Story: for new business goals or new capabilities.
• Feature: for enhancement, modification, or scope adjustment of existing capabilities.
• Issue: for fixing defects and abnormal behavior.

This stage emphasizes clear problem definition and does not enter implementation planning.

2.2 Draft

Draft is responsible for requirement clarification and prototype confirmation. It is the key layer that moves from business language into engineering language.

• Describe core scenarios, boundaries, and exceptional states from the user's perspective.
• Clarify what is in scope and out of scope for this round.
• Provide a discussable prototype or interface outline to align expectations quickly.

The value of this step is to reduce rework: first align on "what to build", then move on to "how to build it".

2.3 Plan / Code / Iteration / Review

Plan: technical decomposition and acceptance checklist stage

• Map requirements into technical tasks and execution order.
• Express success criteria as verifiable acceptance items.
• Clarify risks, dependencies, and rollback strategy.

Code: implementation record stage

• Advance structure design, mocking strategy, test design, and implementation according to the plan.
• Record key decisions and change impact to keep the process traceable.

Iteration: issue convergence stage

• After the first round of test execution, classify issues, fix them, and retest.
• The goal is to converge the main issues into a deliverable state and clearly record remaining items.

Review: delivery confirmation stage

• Analyze differences based on user feedback (expected vs actual).
• Turn confirmed correction items into inputs for the next iteration.

The mapping between Guide and Template is as follows:

Stage	Guide Focus	Template Output
Story / Feature / Issue	Problem definition and scope	St / Ft / Is log
Draft	Requirement clarification, prototype	Df document
Plan	Technical decomposition and acceptance criteria	Pl document
Code	Implementation process and change record	Cd document
Iteration	Test convergence and fix loop	It document
Review	Feedback analysis and correction confirmation	Rv document

3. Human-AI Collaboration Responsibilities

In the A2E collaboration model, humans and AI take different responsibilities:

• Humans are responsible for:
  • Defining goals and trimming scope
  • Making key trade-offs and final acceptance
  • Risk judgment and priority decisions
• AI is responsible for:
  • Drafting documents and structuring content
  • Technical decomposition and implementation progress
  • Test fixes and regression suggestions
• Shared responsibilities:
  • Keeping terminology consistent
  • Keeping documents and implementation in sync
  • Being accountable for explainable and traceable delivery results

This division emphasizes "human-led, AI-accelerated", rather than outsourcing decisions to tools.

4. Quality Gates and Regression Strategy

Quality control has three layers:

1. Documentation gate
   • Every stage must have its corresponding record so goals, implementation, and acceptance remain traceable.
2. Engineering gate
   • Use the existing project quality checks, such as type validation, style checks, builds, and tests, to verify change stability.
3. Delivery gate
   • Completion is not recognized until the user feedback loop is closed in Review.

The regression strategy emphasizes small, fast steps:

• Focus on the critical path first, then extend to edge scenarios.
• Each iteration must clearly state what is resolved, unresolved, or deferred.
• For unresolved issues, keep an actionable next step instead of stopping at description.

5. Lessons Learned and Limitations

Lessons learned:

• A unified Guide + Template can significantly reduce communication cost.
• A document-driven approach breaks complex collaboration into manageable stages.
• The same flow can support product discussion, technical execution, and delivery review at the same time.

Current limitations:

• The system is still in the Experimental stage and continues to evolve in real development work.
• Flow quality depends heavily on input quality: if goals are unclear, later stages amplify the deviation.
• AI can increase speed, but critical decisions still require human control, especially around boundaries and risks.

6. Documentation Directory

• Prompt system entry: Prompt System
• Guide subdirectory: Guides
• Template subdirectory: Templates