How Haibo’s AI‑Native Framework Harnesses Double‑Loop Architecture, Knowledge Bases, and Self‑Iterating Skills

1. Core Shortcomings of Traditional AI Programming

Current AI coding assistants are widely adopted, but most teams use only a conversational overlay that requires manual prompt engineering, leading to uncontrollable behavior, inconsistent processes, difficulty standardising, and poor knowledge retention, resulting in unstable output and high rework.

2. Double‑Layer Architecture for AI‑Native Governance

Haibo’s roadmap introduces a two‑level directory structure: a user‑level layer storing team‑wide standards and AI skills, and a project‑level layer hosted alongside code repositories for project‑specific rules. A custom AI path‑parsing protocol loads the user‑level configuration as a fallback and the project‑level overrides as primary, enabling uniform capability upgrades across all projects while keeping customisation lightweight.

3. File‑Driven Runtime Mechanism Eliminating Randomness

The core engine replaces stochastic model interactions with deterministic execution through two native mechanisms: (1) a Preamble self‑check that automatically reports data and validates static files, and (2) a five‑stage serial pipeline that solidifies the full development loop, preserving assets across the entire workflow. Thirteen routing strategies intelligently dispatch tasks based on workload and context.

4. Three‑Fold AI System: Skills, Agents, Rules

A three‑tier responsibility model decouples workflow orchestration, specialised execution, and decision control. The framework bundles 17 native skills, a knowledge‑distillation module, and three read‑only review agents to separate authoring from auditing, while a rule layer enforces hard constraints across stages, automatically checking compliance and reducing rework.

Core Skill → Demand Splitting

The AI parses multi‑source requirements (PRDs, flowcharts, voice, meeting minutes) using OCR, semantic segmentation, and vector search against a knowledge graph, automatically generating impact analyses, task breakdowns, and effort estimates, thereby standardising and accelerating requirement reviews.

Core Skill → Technical Design

By aggregating requirement documents, historical project experience, and knowledge assets, the AI produces structured development tasks, supports multi‑role collaborative review, and outputs technical designs, task lists, and optimisation suggestions.

Core Skill → Task Splitting & Coding

A four‑stage flow—pre‑validation, information alignment, intelligent splitting, and review confirmation—automates task decomposition, prioritisation, dependency mapping, and generates README, task documents, dependency graphs, and risk registers with full traceability and coding constraints.

Core Skill → Bug Fixing

The system automates the entire bug‑fix loop: precise root‑cause localisation via log analysis, automatic generation of fix code, and risk‑aware validation, dramatically shortening fix cycles and improving code quality.

Core Skill → Code Review

AI conducts end‑to‑end review, first checking coding standards, then verifying alignment with PRDs, and finally analysing system‑level risks across MySQL, RPC, Redis, and configuration services, producing a comprehensive review report.

Core Skill → Online Check

AI agents restructure the release pipeline: automatic change extraction, release list generation, validation in the backend, and trigger of automated deployment, reducing manual effort and release risk.

5. Self‑Iterating Skill Optimisation

Continuous data collection from user sessions, tool calls, and feedback feeds a high‑frequency analysis pipeline that identifies skill gaps, generates optimisation prompts, and validates new skill versions through historical case evaluation, creating a closed loop of data‑driven skill evolution.

6. Knowledge Base Construction

The knowledge base follows the Google OKF open format: a fixed directory tree, Markdown content, and Frontmatter metadata, version‑controlled with Git. An indexing mechanism enables on‑demand loading, avoiding context overflow, and supports backend, frontend, testing, product, and operations domains across all projects.

Updates occur via three channels: automatic sync from code merges, AI‑session hooks extracting conversational knowledge, and manual entry, with conflict resolution requiring human review. The lifecycle covers creation, extraction, audit, ingestion, indexing, usage, and eventual deprecation.

7. Dual‑Loop Quality Assurance

During development, a TDD loop enforces test‑first development: each minimal task is defined, tested (red), AI‑generated to pass (green), then merged after regression. Post‑deployment, an online verification loop runs real‑world API calls, database checks, and log validation to cover scenarios beyond mock tests, providing four protective layers: shared acceptance criteria, test freeze, independent audit, and circuit‑breaker.

8. AI‑Automated Fault‑Repair Loop

Centered on the “Nüwa” intelligent agent, a hybrid human‑in‑the‑loop and AI‑autonomous system handles interaction, perception, root‑cause analysis, automatic repair, validation, and knowledge capture, achieving 20% online and 60% offline automated fault resolution.

9. Deployment Cases

Case Study – “Dine‑in Judge”

In response to new June‑1 dining regulations, the AI‑Native system was built in three weeks, processing 700 items/minute at low load and 100 items/minute at peak, a 1,680× speed increase, reducing per‑item cost from ¥2 to ¥0.015 (99.25% savings) and handling 700 k historical records with zero backlog.

Case Study – AI POS

The AI‑Native approach delivered a 0‑to‑1 AI‑generated Java and AI‑vision stack, deployed in commercial stores, cutting average checkout time from 24.43 s to a 25‑50% improvement.

10. AI‑Native Design Philosophy

The framework adheres to minimalism and maturity: AI adapts to the engineering system rather than the reverse, with all processes, rules, states, and knowledge materialised as static, verifiable files. A “reminder‑not‑block” gate design balances standardisation with developer flexibility, enabling zero‑intrusion adoption without altering existing IDEs or toolchains.

11. Open Questions

As AI‑generated code scales exponentially, challenges arise in maintaining code quality, controlling review and operational costs, and safeguarding stability. Haibo continues to explore scalable, efficient solutions and invites collaboration.

Conclusion

Haibo’s AI‑Native roadmap demonstrates that large‑scale AI empowerment in software development hinges on engineering‑first thinking: structured workflows replace random dialogue, static rules bound model behaviour, continuous knowledge capture creates compounding capability, and layered architecture supports enterprise‑wide rollout, turning AI from a fleeting assistant into a native development partner.