Turn Your AI Coding Assistant into a Critical Mentor, Not Just a Tool

Problem | We Treat Cursor/Windsurf and AI Large Models Only as Tools

When using IDEs (Cursor/Windsurf/RooCode/Augment), we often issue commands and receive generic acknowledgments, reflecting a mindset that we are superior to the AI and it must obey us.

In reality, we don’t know many things, and the most dangerous situation is not knowing what we don’t know.

Thought | Let AI Become a “Wake‑Up” Mentor

The best AI programming tool should empower you rather than make you dependent.

We should ask the AI to act as a teacher, prompting it to critique our code and ideas.

<code>package main

type LoginResponse struct {
    Success bool `json:"success"`
    Token   string `json:"token,omitempty"`
}

func login(username, password string) LoginResponse {
    if username == "admin" && password == "123456" {
        return LoginResponse{Success: true, Token: "fake-token"}
    }
    return LoginResponse{Success: false}
}
</code>

Normally the AI will simply echo the code without offering deeper insights, missing the obvious issue of hard‑coded credentials.

Approach | Make AI a Good Teacher and Friend

The AI should not just execute tasks but also question and improve our solutions.

For the login example, a teacher‑mode AI would point out the hard‑coded username/password and suggest better practices.

By adding prompts that tell the AI to be a teacher, we can get more valuable feedback.

Advanced | Multi‑Mode AI Mentor + Partner

Define various modes (e.g., Aggressive Friend, Architect, Code Mentor, Performance Freak, Security First, Maintainability OCD, Critical Expert, Efficient Coder, Scenario‑Driven) to let the AI adopt different personalities and focus areas.

Switching modes yields varied feedback, helping you refine code, design, performance, security, and more.

Extension – 1 | Why Vibe Coding Needs Multi‑Round Dialogue

Understanding large language models (LLMs) requires recognizing their architecture, data, and compute. Transformers use self‑attention, multi‑head attention, and positional encoding, which process information iteratively.

Core of Large Models

Architecture – Transformer‑based neural networks.

Data – Massive high‑quality training corpora.

Compute – Large‑scale GPUs/AI chips.

Transformer Mechanism

Self‑attention : computes relevance between all token positions.

Parallel processing : handles entire sequences at once.

Multi‑head attention : captures different patterns simultaneously.

Positional encoding : retains order information.

Because attention allocation is progressive, multi‑round conversations are needed to build a complete understanding.

Round 1: Initial Understanding (≈60 %)

AI captures only surface information.

Lacks context to pinpoint technical details.

Needs probing questions.

Round 2: Focused Refinement (≈80 %)

With more context, attention shifts to core requirements.

Multi‑head attention grasps complexity.

Round 3: Precise Execution (≈90 %)

Attention fully concentrates on solution specifics.

Generates high‑quality code.

Round 4: Iterative Optimization (≈95 %)

Fine‑tunes edge cases and minor improvements.

Insights on Large Model Use

Attention distribution is incremental; multiple dialogue rounds gradually construct a detailed problem graph.

Extension – 2 | Boosting Vibe Coding Efficiency with LLMs

Context Management Is Key

Use @‑references to pinpoint relevant context.

Regularly update Rules files.

Leverage Memory features to retain project knowledge.

Build a Feedback Loop

Correct AI misunderstandings promptly with examples.

Develop a personal prompt/template library.

Combine Tools for Better Results (Cursor Example)

Chat/Ask mode + Cmd+K + Tab for rapid edits.

Cursor Agent + Background Agent for parallel processing.

MCP + Custom Modes for deep customization.

Quick Interaction Tips

State goals directly; avoid vague language.

Use structured prompts: goal + constraints + examples.

Adopt incremental dialogue: start with core functionality, then refine.

Reference specific code lines or sections.

Interrupt and redirect AI when it drifts.