Fred Brooks once said: “Complexity doesn’t disappear, it only transfers.”
While AI can handle much of the complexity for us, due to the inherent complexity of software development and AI’s intrinsic limitations (context forgetting, hallucinations, sycophancy, etc.), completing large projects solo still faces many challenges.
Have you encountered these problems when developing with AI: thoughts tangled like a mess in your head, unable to track which designs are complete and which are yet to be implemented; when modifying old designs, just getting the AI to locate the corresponding code consumes most of your tokens?
This project implements an AI-oriented engineering protocol that solves the problem of individuals being unable to manage large projects through the following mechanisms:
SSOT (Single Source of Truth) + Ledger Index: Compiles “high-value human intuition” into versionable, referenceable, and verifiable truth artifacts, eliminating information gaps and context forgetting.
Knife Slicing Algorithm: Converts complex requirements into a queue of atomic change packages that AI can process, solving cognitive bandwidth constraints.
Layered Quality Gates (G0-G6) + Evidence Closure: Replaces “self-declared completion” with executable checks, preventing hallucinations and false convergence.
Role Isolation and Handoff Contracts: Avoids the self-approval problem where AI writes and validates its own work.
Recommended Usage: Assume you are maintaining a continuously evolving SSOT document. The traditional approach requires you to remember which designs have been implemented and which haven’t, then tell the AI: “I’ve added XX, please help me implement it.” But as the project continues to evolve, developers inevitably misremember or forget things, causing implementation to drift from design. This project uses ledger indexing + change package traceability mechanisms to let the system automatically track the complete chain of “design → implementation → verification → archival,” so developers only need to focus on intent input without manually maintaining state.
This project is heavy because complexity doesn’t disappear, it only transfers. I personally believe that if you want to achieve high automation, heavy processes are unavoidable. It’s suitable for power users with sufficient token budgets, quality requirements for development, and pursuit of full automation, best used with AI programming software that supports parallel sub-agents/parallel windows.
Personal Experience: By using the delivery skill to submit a requirement, I can run the entire workflow with a single command in Codex without manual confirmation. Here is my commonly used prompt:
Your Action Guidelines:
As the main agent, do not execute any node tasks (e.g., test owner, proposal author). All node tasks should be delegated to sub-agents. Both the main agent and sub-agents must explicitly choose a real-world human master who excels at global understanding of the current task to roleplay at the beginning of each task. Except for coder and test owner which use the gpt-5.2-codex xhigh model, all others use the gpt-5.2 xhigh model. All decisions are your responsibility; you need to have a global and long-term perspective.
All sub-agent outputs: If modifying the same file, modify it directly. If outputs are independent, output them to the change package and clearly mark which iteration it is, to avoid confusion when distinguishing multiple challenges.
If there are other changes in progress simultaneously, please don’t mix them up. Always declare the current change package. If starting a change from scratch, explicitly instruct the proposal author to create the change package. If a change is already in progress, check the current progress (query key artifacts) and then continue.
The process only ends after running the archiver. Do not stop midway to ask the user.
As the main agent, you must strictly control quality. No cutting corners or premature completion. The token budget is very sufficient, and the current environment has context compression mechanisms, so feel free to go all out and maintain strict control.
Important: If you’re working in Codex CLI, don’t be impatient when calling sub-agents. Wait for sub-agents to fully complete their tasks before sending messages, rather than frequently asking if they’re done or requesting progress reports.
Github URL: https://github.com/Darkbluelr/dev-playbooks
Design Book: https://github.com/Darkbluelr/dev-playbooks/blob/master/docs/ai-native-workflow.md
Top comments (0)