1 of 58

Vibe Coding

Chat, Coding Agents and Work

Industrial AI Lab, KAIST

1

2 of 58

Vibe Coding

Describe what you want in natural language; the AI writes the code
Role shift

Goals and the big picture instead of implementation details

Wider access

Software without professional programming training

Conversational iteration

Prompt → check the result → revise the prompt

2

3 of 58

Three Ways to Vibe Code

3

	Who Runs the Code	What It Remembers	What It Delivers
Chat (Gemini, ChatGPT, Claude)	Human	This conversation only	Code snippets
Coding Agent CLI (Claude Code, Codex, Gemini CLI)	Agent	Project rule files	Working code and results
Work (Claude Cowork, ChatGPT Work)	Agent	Projects and memory	Reports and other work products

4 of 58

MNIST database

4

5 of 58

Problem: Handwritten Digit Classification with MNIST

Task

Build an Artificial Neural Network (ANN) that classifies handwritten digits from 0 to 9 using the MNIST dataset.

Requirements

Use the external MNIST dataset
Design the ANN architecture
Train and validate the model
Evaluate performance on the test dataset
Report the model architecture and test results

Input: 28 × 28 grayscale handwritten digit image
Output: Predicted digit class (0–9)

5

6 of 58

Chat

6

7 of 58

Human Runs the Code

Chat: the LLM writes the code; the human runs it, checks it and asks again

7

Chat

Q

Human

LLM

Code

Human

runs in Colab, edits, asks again

8 of 58

Prompt

“Write a neural network code that classifies MNIST”

8

9 of 58

Prompt

“Write a neural network code that classifies MNIST”

9

10 of 58

Prompt

“Write a neural network code that classifies MNIST”

10

11 of 58

Strengths and Limits

Strengths

Easy to start
Human reviews the output
Human controls where the code runs

Limits

Cannot access local files directly
Limited continuity across sessions

11

The AI writes the code; the human runs it and checks the result

12 of 58

Coding Agent CLI

12

13 of 58

Agent = LLM + Tools + Loop

LLM: reads and writes text and code

13

Goal

Human

Large Language Model

Tools

Files

>_

Terminal

Python

Result

Check

Report

Revise

14 of 58

Agent = LLM + Tools + Loop

LLM: reads and writes text and code
Tools: files, terminal, Python

14

Goal

Human

Large Language Model

Tools

Files

>_

Terminal

Python

Result

Check

Report

Revise

15 of 58

Agent = LLM + Tools + Loop

LLM: reads and writes text and code
Tools: files, terminal, Python
Loop: check the result, then revise

15

Goal

Human

Large Language Model

Tools

Files

>_

Terminal

Python

Result

Check

Report

Revise

16 of 58

Tools You Can Use Today

Same loop, different companies

16

Claude Code

Anthropic

Codex

OpenAI

Gemini CLI

Google

Claude Code CLI

17 of 58

Installing Claude Code

Desktop app for Windows: download it, then run the installer
Needs a paid Claude plan (Pro, Max, Team, Enterprise) or an API account

17

① Download for Windows

② Run the installer

18 of 58

Other Ways to Run Claude Code

Terminal
VS Code
This lecture uses the desktop app

18

① Terminal (CLI)

② Inside the editor (Extension)

19 of 58

Folder to Work On

One folder per project

19

① New folder: SciML HW

20 of 58

Choosing the Folder

A local session works inside the folder you choose
Claude reads and edits the files in that folder

20

① Select it

② Select Folder

21 of 58

Session Screen

Four things to check before the prompt

21

① Folder

② Prompt

③ Permission mode

④ Model and effort

22 of 58

Choosing the Model

Bigger models for harder tasks

a small one is faster for easy ones

22

① Used in this demo

② Older models

23 of 58

Setting the Effort

How long Claude thinks before it acts: from Low to Max
Default Medium for Opus 5.5

set to High for this homework

23

① Now High

② Slider

24 of 58

Permission Modes

How much Claude may do without asking you

24

Default: Claude decides, risky actions blocked

Asks before every change

Edits files without asking

Plans first, changes nothing

Leave off

25 of 58

CLAUDE.md / AGENTS.md

Project rules in one Markdown file, read in full at every start
CLAUDE.md for Claude Code; AGENTS.md for Codex, Gemini CLI, Cursor

25

Goal

Human

Large Language Model

Tools

Files

>_

Terminal

Python

Result

Check

Report

Revise

CLAUDE.md

every start

Never download data and never modify anything in `data/`.

Flatten → Linear(784, hidden) → ReLU → Linear(hidden, 10)

… do not silently reduce epochs, data, or configurations.

demo/mnist-hw/CLAUDE.md

26 of 58

First Prompt

Prompt: “Creat CLAUDE.md file in the current folder and write breifly about SciML.”

26

① This folder

② The task

③ Opus 5.5, High

27 of 58

File Written by Claude

CLAUDE.md appears in the folder

27

28 of 58

Inside CLAUDE.md

Markdown

File extension

28

29 of 58

Skill

A saved procedure for a task you repeat, called by name
Until it is called, only its name and description are in the context

29

Goal

Human

Goal

Human

LLM

Without

a skill

With

a skill

1. Write train.py (hidden size, epochs, lr, seed)

2. Run each setting, one subagent each

3. Write summary.md and comparison.png

4. Reply with the table and the file paths

typed into the prompt every time

/mnist-experiment

Compare hidden sizes 100, 256, and 512 for our MNIST homework model.

one command, typed once

SKILL.md

the same 4 steps,

written once

loaded

30 of 58

SKILL.md

The demo’s skill: plain Markdown, a short header and the steps

30

always in context

loaded when called

31 of 58

Creating a Skill

Ask Claude to write the skill for you

31

Where to save it

Name, when to load it

The procedure

32 of 58

Skill Folder

One folder per skill, with its SKILL.md inside
Start a new session so Claude finds the new skill

32

① Your .claude\skills

② New skill

33 of 58

‘homework-solver’ Skill

Same shape as the MNIST skill

33

Always in context

Loaded when used

34 of 58

Using the Skill

One sentence
The request matched the skill’s description

Claude loaded it

34

① Homework notebook

② The prompt

35 of 58

Claude at Work

The agent loop

run, check, revise

35

① A command failed

② It changes its approach

③ It finds the nine red problems

36 of 58

Claude at Work

The agent loop

run, check, revise

36

④ Runs on a copy first

⑤ Finds its own mistake

37 of 58

A Solution in the Notebook

Each problem gets a Markdown cell and a code cell

37

① Explanation

② Equations

38 of 58

A Solution in the Notebook

38

39 of 58

Checking the Results

Claude’s own check of every problem, with what it fixed
“Verified” is Claude’s check, not a grade: you still review it

39

① A real error, found and fixed

40 of 58

Tips for Your First Tasks

Be specific

The file, the constraints, an example

Give a way to check

An expected number, a test, a plot

One task per session

A new session for an unrelated task

40

41 of 58

Subagent

A helper agent with its own context; only a summary comes back
.claude/agents/<name>.md, with its own tools and model

41

Goal

Human

Large Language Model

Tools

Files

>_

Terminal

Python

Result

Check

Report

Revise

mnist-trainer

h = 100

97.48 %

mnist-trainer

h = 256

97.58 %

mnist-trainer

h = 512

97.84 %

task out,

summary back

42 of 58

MCP (Model Context Protocol)

One open standard that connects outside services to an agent
A connected server adds its tools; used by ChatGPT, Gemini, Cursor, VS Code

42

Goal

Human

Large Language Model

Tools

Files

>_

Terminal

Python

Result

Check

Report

Revise

Google Drive

Notion

GitHub

MCP servers

MCP

43 of 58

Plugin

Skills, subagents, hooks and MCP servers installed as one unit
Shared through a marketplace with /plugin; CLAUDE.md stays per project

43

Goal

Human

Large Language Model

Tools

Files

>_

Terminal

Python

Result

Check

Report

Revise

Plugin: ml-homework (example)

CLAUDE.md

Skill

MCP server

Marketplace

User

/plugin

44 of 58

Demo: MNIST with a Coding Agent

One instruction, no human input: code, training, evaluation, comparison file, report

44

45 of 58

Where It Runs

My computer

My GPU, no time limit

Cloud

Runs with the laptop closed
No GPU, time limit on each command

45

46 of 58

Agent: Strengths and Limits

Strengths

Can handle multi-step tasks
Can write, run, and fix code
Can work with files and tools
Reduces manual work

Limits

Needs clear instructions
Results still need human checking

46

Agents can run experiments, evaluate results, and improve the model automatically

47 of 58

Work

47

48 of 58

Project, Memory and Folder Access

48

Concept	What It Does	In the MNIST Homework
Project	Folders, instructions and material grouped per course or assignment	One project per course
Memory	Context kept across sessions, per project	Course rules and earlier results remembered
Folder access	Reads a chosen local folder and saves results there	Reads the results folder of the coding agent run

49 of 58

Work: Agents for the Whole Task

Not only code: the whole task
In an app, no terminal
Claude Cowork, ChatGPT Work

49

Project

50 of 58

Work: Strengths and Limits

Strengths

Familiar to non-developers
Strong on documents and slides
Can handle multi-step tasks
Keeps context across tasks

Limits

Less control over the runtime environment
Limited control over compute resources
Not designed for heavy GPU training

50

Human sets the goal, AI does the work

51 of 58

Which One When

Chat

Learning concepts, writing short functions, understanding errors

Coding Agent

Running experiments, comparing results, fixing code

Work

Creating reports and slides, carrying context across assignments

51

Chat

Coding Agent

Work

understand

experiment

report

52 of 58

Which One When

Chat

Learning concepts, writing short functions, understanding errors

Coding Agent

Running experiments, comparing results, fixing code

Work

Creating reports and slides, carrying context across assignments

52

Chat

Coding Agent

Work

understand

experiment

report

Choose the right tool for your goal and cost

53 of 58

Models and Effort

Model: Brain
Effort: how long it thinks before answering

Low, medium, high, xhigh, max

53

Source: https://artificialanalysis.ai/models

54 of 58

Raising the Effort

Opus 5.5, Low to Max
Low to High: large gain
Beyond High: flat
Max below Xhigh

54

Source: https://www.anthropic.com/claude-opus-5-5

55 of 58

Higher Is Not Always Better

FrontierCode v1.1
Default Med above High and Xhigh

55

Source: https://www.anthropic.com/claude-opus-5-5

56 of 58

Time per Task

Opus 5.5, Low to Max

Time per task: 1.4 to 13.0 min

56

Source: https://artificialanalysis.ai/models

57 of 58

Choosing the Effort

Easy work at high effort: only slower and costlier
Plan with a strong setting, execute with a light one
Start at the default (medium); raise it only when stuck

57

58 of 58

Thank you

58