xop01/ai_goodpractice

ai_goodpractice is a hand-written guide (not an AI-generated one) by the developer xop01, sharing how they use AI to act as a manager of coding work rather than doing the coding themselves.

What it is

A lengthy, opinionated guide based on the author's daily experience (over 8 hours per day using AI for many months). They claim agentic workflows — where an AI assistant writes, tests, and fixes code while a human directs it — let you 10–20x your output and complete ambitious software projects in months instead of years.

The author wrote the guide manually on purpose: they believe AI can't express their own thoughts and lifestyle as well as they can. They invite feedback and discussion on their X (Twitter) account.

The core idea

Instead of being a programmer who types code, you become a "leader/manager" who:

• Envisions the software and how it should be built
• Formulates that vision into detailed prompts the AI can execute
• Reviews and steers the AI's work, like a manager directing a team

The author says they have not written a single line of code manually since adopting this workflow.

Who it's for

The guide claims to be useful for complete beginners as well as experienced coders — anyone who wants to get real, sustained work out of AI subscriptions.

The guide's structure: 3 steps

Step 1 — Setup

• Pick an AI-integrated IDE (Cursor, VS Code, Claude Code, ChatGPT, etc.)
• Treat the AI as a capable partner who still needs clear instructions — think about whether your prompt would make sense to a smart person with no extra context
• Set up the environment with rules, skills, and MCPs (small tools that connect AI to other software). The author calls this the most underrated step.

Key practice: build your rules through the AI itself. Instead of writing rules manually, prompt the AI to analyze your code style and generate a rule for the repo. The author gives example prompts and stresses being detailed — e.g., "analyze the code style in depth and apply it as a rule," plus special rules for tests, helpers, and simplicity.

Step 2 — Avoiding "slop" (low-quality AI output)

The author warns that much "AI slop" comes from weak foundations and careless prompting. Their recommendations:

• Prefer making your own rules/skills/MCPs over downloading ones from the web, since off-the-shelf ones often don't fit your workflow
• The author's own MCP: project-progress-mcp — a tool (written in Zig, the author says they'd never used Zig before) that tracks project progress so the AI keeps a documented roadmap and doesn't drift from original goals
• Documentation is critical: keep lightweight docs about key features and infrastructure so the AI can quickly understand the codebase without re-reading whole files. But keep docs updated, or they become stale and misleading
• Tests are the most important step: a proper test suite lets the AI validate its own work and work autonomously for long stretches without a human checking everything
• For big tasks: use your IDE's "plan mode," write a very detailed plan, specify that every major change needs a test case, tell the AI time isn't a constraint ("the plan has no time limit"), then review the whole plan carefully before starting
• Audit at the end: after the plan is done, prompt the AI to confirm everything was implemented for "longevity, scalability and stability," referencing the original plan

Step 3 — Making AI improve your life

The author lists 7 principles. The guide elaborates on the first two:

1. Use your time to maximize prompting — match your prompts to how long the AI will take. If you're leaving to cook dinner, launch a prompt big enough to keep the AI busy while you're away. This lets you "work" many hours while resting in between — but the author warns against constantly checking the AI's progress ("moderation is key").
2. Learn the stack or subject you work with — AI is a tool to speed you up, not a replacement for understanding. To direct the AI well and catch mistakes, you need at least a basic grasp of what it's working on.

The guide lists but does not detail the remaining five: know when to walk away, trust your partner, translating your thoughts, multi-tasking, and use the right model for the right task.

Example prompts from the guide

• "I need you to analyze the code style from (folder path/url) in depth and apply it as a rule for this repo"
• "I need you to create a skill that agents will use to work on MySQL... go in-depth with MySQL possibilities"
• "This project will be used to (purpose)... I need recommendations on what kind of stack would be the best... based on my needs: longevity, stability, scalability, performance"

Bottom line

The project is essentially an opinionated essay, not code. It argues that with disciplined setup, detailed prompting, documentation, and tests, one person can get AI to build serious software — and that the same habits translate into a better work-life balance, letting you work fewer focused hours while the AI carries the load.