LLM Coding Workflow Guide: A Practical Guide to Software Development with Dual-Model Collaboration

Original Author/Maintainer: Jaksa576 Source Platform: GitHub Original Link: https://github.com/Jaksa576/llm-coding-workflow-guide Publication Date: May 24, 2026

Core Idea: Achieve efficient LLM-assisted software development by separating planning and execution contexts using ChatGPT (as a planning partner) and Codex (as a coding agent). The core concept is "Controlled Autonomy"—AI takes on repetitive planning, coding, documentation, and QA support tasks, while users maintain control over product direction, judgment, and final approval.

With the improvement of large language model capabilities, core issues have emerged when developers try AI-assisted development: a single model struggles to simultaneously handle high-level planning (product thinking, architectural decisions, trade-offs) and specific code implementation (precise code knowledge, repository awareness, tool operation).

Core Insight: Separating planning and execution contexts into different models yields better results—ChatGPT handles product thinking, architectural decisions, documentation writing, and quality assurance; Codex focuses on repository-aware implementation, validation, and submission.

Core Concept "Controlled Autonomy": AI undertakes planning, coding, documentation, and QA support; users control product direction, judgment, and final approval, balancing AI capabilities and risks.

Responsibility Separation:

• User: Owns ideas, roadmap judgment, QA judgment, and merge decisions
• ChatGPT: Converts rough intentions into plans, documents, QA classifications, and handovers to coding agents
• Codex: Implements for local repositories, performs validation work, updates documents, submits, and reports
• GitHub Repository: Serves as the source of truth

Applicable Scenarios: Hobbyists/independent developers building real software while maintaining control; not applicable to one-off small scripts, pure no-code applications, or high-risk production systems.

Importance of Context Separation:

1. Token Efficiency: Only send task-specific context to Codex, saving tokens and reducing latency
2. Context Hygiene: Prevent coding agents from being overwhelmed by outdated/irrelevant information

14-Stage Workflow: From repository creation (Stage 0) to continuous implementation loops (Stages 9-14), covering repository configuration, ChatGPT project creation, document generation, handover for implementation, and other links

Document Strategy: Source of Truth Hierarchy (HTML Guide, Concise Workflow Primer, Project Instructions, Repository Documentation)

Handover Design: Includes repository status, task scope, acceptance criteria, validation strategy, documentation requirements, and stop conditions; keeps it concise with only necessary information.

Default Environment: Windows + PowerShell (reduces variables, suitable for most developers)

Tool Assumptions:

• ChatGPT: Planning (product thinking, architectural decisions, documentation writing, QA decisions)
• Codex: Implementation (repository-aware coding, validation, document updates, submission)
• GitHub: Source of truth and collaboration platform
• Local Repository: Development environment

Tools can be replaced, but the core is the architecture of separating planning and execution.

Main Trade-off: Setup overhead (may not be worth it for one-off projects)

Applicable Scenarios: Hobbyists/independent developers building real software and needing to maintain control

Non-Applicable Scenarios:

• One-off small scripts
• Pure no-code application generation
• High-risk production systems requiring professional engineering review

Even without using ChatGPT/Codex, the core ideas still have universal value:

1. Context Separation Principle: Reduce cognitive load and improve output quality
2. Documentation as Code: Maintain in sync with code instead of supplementing afterward
3. Controlled Autonomy Framework: A middle ground for human-AI collaboration, balancing AI capabilities and supervision
4. Handover Design Pattern: Reference for multi-agent/human-AI collaboration, clarifying the minimal necessary content for information transfer

The core contributions of the guide lie in the two key concepts of "separation of planning and execution" and "controlled autonomy", providing a sustainable and scalable paradigm for LLM-assisted software development.

For teams/individuals exploring AI-integrated development, it offers a validated reference implementation, which is worth adjusting according to their own needs.