# AI

# What i've learned in a month of using Claude

1. Cost Optimize 
    1. RTK
    2. Cavemen
2. 9Router 
    1. Free LLM Provider
3. Opus is smarter > Codex > Others
4. Swarm > Workspace

What Model for What?

Only use Opus for Superpowers Planning > Claude  
Coding > Codex > GLM  
Scout > Free > Git Nexus > Living Atlas > Repo Search

# How to run Tauri Desktop App

commands I ran to take this Windows host from "no Rust" to "running Tauri desktop app":

1\. Detect prerequisites  
\# Existing C++ toolchain?  
$vsw = "${env:ProgramFiles(x86)}\\Microsoft Visual Studio\\Installer\\vswhere.exe"  
& $vsw -products \* -requires Microsoft.VisualStudio.Component.VC.Tools.x86.x64 -property installationPath

**\# Rust/cargo/linker present?**  
Get-Command rustup -ErrorAction SilentlyContinue  
Test-Path "$env:USERPROFILE\\.cargo\\bin\\cargo.exe"  
Get-Command link.exe, cl.exe -ErrorAction SilentlyContinue  
2\. Install Rust  
winget install --id Rustlang.Rustup -e `  
\--accept-package-agreements --accept-source-agreements --silent  
3\. Activate MSVC toolchain  
$env:PATH = "$env:USERPROFILE\\.cargo\\bin;$env:PATH"  
& "$env:USERPROFILE\\.cargo\\bin\\rustup.exe" default stable-msvc  
& "$env:USERPROFILE\\.cargo\\bin\\rustc.exe" --version # verify  
& "$env:USERPROFILE\\.cargo\\bin\\cargo.exe" --version # verify  
4\. Copy project off WSL share to local disk  
New-Item -ItemType Directory -Path C:\\dev # only if missing  
robocopy "\\\\wsl.localhost\\Ubuntu-22.04\\home\\son\\r3nk\\workbrain" `  
 "C:\\dev\\workbrain" /E `  
 /XD node\_modules target .next out .turbo `  
 /XF "\*.log" /NFL /NDL /NP /R:1 /W:1  
5\. Check port collision  
Get-NetTCPConnection -LocalPort 3000 -State Listen  
Get-NetTCPConnection -LocalPort 3001 -State Listen  
\# Found WSL relay holding port 3000 → had to move our Next dev to 3010  
6\. Rewire Tauri devUrl to port 3010  
Edited src-tauri/tauri.conf.json:

\- "beforeDevCommand": "pnpm dev",  
\- "devUrl": "http://localhost:3000",  
\+ "beforeDevCommand": "pnpm dev -p 3010",  
\+ "devUrl": "http://localhost:3010",  
7\. Install JS deps  
Set-Location C:\\dev\\workbrain  
pnpm install  
8\. Launch the app (background, with logs)  
New-Item -ItemType Directory C:\\dev\\workbrain\\logs # if missing  
$env:PATH = "$env:USERPROFILE\\.cargo\\bin;$env:PATH"  
Start-Process -FilePath "pnpm.cmd" -ArgumentList "tauri","dev" `  
 -WorkingDirectory "C:\\dev\\workbrain" `  
 -RedirectStandardOutput "C:\\dev\\workbrain\\logs\\tauri-dev.out.log" `  
 -RedirectStandardError "C:\\dev\\workbrain\\logs\\tauri-dev.err.log" `  
 -PassThru  
9\. Monitor the build / confirm window opened  
**\# Tail build progress**  
Get-Content C:\\dev\\workbrain\\logs\\tauri-dev.err.log -Tail 30

**\# Confirm the desktop window process**  
Get-Process WorkBrain | Select-Object Id, ProcessName, StartTime, MainWindowTitle  
To stop later  
Get-Process workbrain, pnpm, node, cargo, rustc -ErrorAction SilentlyContinue | Stop-Process  
To restart (cargo cache is warm — fast)  
cd C:\\dev\\workbrain  
pnpm tauri:dev

stay in the current window and prepend cargo for just this session:

set PATH=%USERPROFILE%\\.cargo\\bin;%PATH%

# Những gì claude đã giúp giải quyết

Bình thường nếu mà mần config hosting, server, rồi npm, security check sẽ tốn cả tuần. nếu mà ngồi mần tay trên proxmox hetzner thì tốn biết bao nhiêu là thời gian. giờ claude nó làm trong một nốt nhạc trong 3 tiếng.

# What Local AI can do?

First of all, we need to understand the overalll of all tools, models, and services:

Programming:

- Claude is the king -> Critical Planner (Expensive )
- Codex is Queen -> Broad Planner ( Expensive )
- GLM / Kimi K3 -> Worker
- Qwen, Gemma -> Free

Image Generation:

- Google Gemini
- GPT

Video Generation:

- Seedance
- Veo

CLI with good system prompts:

- Claude
- Pi
- Gemini

Top Agentic Apps:

- Hermes Agent

Local LLM

- Fable gemma
- Fable qwen
- Bonsai convert from big model to small

What can local llm can do right now:

1. does it smart enough to follow exactly the instruction of my language?
2. does it able to do long tasks?
3. how fast token generation it can do on our a 24 GB VRAM GPU 3090?
4. what dataset do we have and public have? old vs new?

# How can i change from one image to relevant image?

Which is good image generation service/provider?

Is it google? google workflow? stitch? banana? seedance? gpt image? claude?