My AGENTS.md

[corz]

I've been doing a lot of AutoIt coding recently, catching up on finishing off some apps, polishing GUIs, that sort of thing. Of course, I have no time for grunt work or messing with GUI crap, I just ask an agent to do it.

I'm getting great results, so I thought I'd share my AGENTS.md.

You will want to edit it, but it will give you a good starting point for getting agents to write code the way you like it, keep good backups, and much more. Some of my instructions won't apply (e.g. I run Codex in its own user account, so it can't use "rg" and I can't be bothered installing it for my Agent user, and so on). My needs are specific, and hopefully this shows you how you can get your specific needs met, too.

All agents will read this file (I keep hardlinked versions in all my project folders, as they will often "forget" to look in the parent folder for instructions) at the start of any thread, and I've barely edited it for weeks now, so it must be working ok!

Have fun vibing AutoIt!

AGENTS.md

Edited September 21 by corz

[mLipok]

my global AGENTS.md is about 15k
and for one of my new project 54k
But I even have specific rules for coloring and GUI positioning.

[mLipok]

What kind of AI you are using ?
I currently use ChatGPT  for simple chat and with CODEX variant.

 

[corz]

Codex is good, dunno about "Variants", I just have Codex (though it can switch to chatty ChatGPT mode in the top-left corner, I notice. Not something I would ever click; I can use the web chat if I want to know which hot sauce is required).

Codex is pretty good, with a great understanding of AutoIt code at Sol Light and Medium (light can handle most jobs, medium is best for more complicated stuff), and I use it often (I have a free offer Plus plan right now and it's been useful) for coding tasks with good success. It helped me with a pretty huge translation job recently and the results were as good as I could have coded myself, and in about 1/1000th of the time.

I mainly use OpenCode, currently with Muse Spark 1.3 Contributor on High (XHigh sucks, and burns more tokens, though you still won't hit your OpenCode Go limit). The "Contributor" tier means they openly train on your data, but as everyone is training on your data I don't see an issue. When Muse Spark 1.4 comes out, I want it trained on how *I* like to work, and LOTS of AutoIt code.

Subscribe to OpenCode Go, is the best advice I can give in this regard. For $10/mo you basically have unlimited usage of a shit-ton of top models; Muse Spark 1.3, GLM 5.3 (what huggingface used to combat the OpenAI attack), the latest DeepSeek, Qwen, Mimo, and so on. Some models allow more tokens than others for your subscription cost, Muse Spark's allowance is MASSIVE. Basically, with an OpenCode Go subscription you get 6x compute cost, so $60/mo for $10, bulk buying, redundant compute, etc., makes it viable for them, cheap for us!

I use the CLI version of OpenCode when working with local (usually Qwen, good at code) models; long, slow tasks. I use the desktop version for back-and-forth vibe coding stuff with Muse.

Note: that last section of the AGENTS.md makes a *MASSIVE* difference when working with Muse Spark. Previously, reading its responses was like trying to parse machine code. Now it even replies to cultural references like "This is the way" with, wait for it: "This is the way"

Edited September 21 by corz

[corz]

1 hour ago, mLipok said:

my global AGENTS.md is about 15k
and for one of my new project 54k
But I even have specific rules for coloring and GUI positioning.

That sounds too big to me. 

I have found that the more you add, the more is lost. Also, that shit needs to get sent with EVERY prompt.

Hey, tell you who is good at trimming an AGENTS.md down so that it will actually be read and followed: AI!

Edited September 21 by corz

[argumentum]

..have you tried anything locally ( your GPU ) ?

[corz]

7 minutes ago, argumentum said:

..have you tried anything locally ( your GPU ) ?

Hah! You missed my recent post!

YES! Lots. I use Qwen 3.6/3.8 a fair bit. It's good at code and long agentic stuff, and spawning a shit-ton of sub-agents! The 35B/3B models are shaky, but the 27B dense models are actually shit-hot at code. I've ran head-to head challenges and then ran the results through frontier models for analysis and every time the 27B model wins, better architecture, just better code, thinks more like we do, keeping edge cases in mind. Muse is also really good at spotting edge cases, but that's not local!

Nothing fancy, I just have llama.cpp running (it just uses the models from my ComfyUI LLavacheckpoints dir) and OpenCode uses that. I have a batch file to choose what model I want to run right now:

@echo off
setlocal enabledelayedexpansion

:: ----------------------------------------------------
:: Configuration Pathways
:: ----------------------------------------------------
set "SERVER_EXE=I:\AI\llama.cpp\build\bin\Release\llama-server.exe"
set "MODEL_DIR=I:\AI\models\LLavacheckpoints"

:: Verify directory exists
if not exist "%MODEL_DIR%" (
    echo [ERROR] Model directory not found: %MODEL_DIR%
    pause
    exit /b
)

:MENU
cls
echo ====================================================
echo                 Available GGUF Models
echo ====================================================
echo.

:: Loop through directory and index GGUF files
set "count=0"
for %%F in ("%MODEL_DIR%\*.gguf") do (
    set "fname=%%~nxF"
    
    :: Skip files containing "mmproj" (case-insensitive check)
    if "!fname:mmproj=!"=="!fname!" (
        set /a count+=1
        set "model[!count!]=!fname!"
        set "model_path[!count!]=%%~fF"
        echo  [!count!] !fname!
    )
)

if %count%==0 (
    echo [ERROR] No eligible .gguf models found in %MODEL_DIR%
    pause
    exit /b
)

echo.
echo ====================================================
echo.

:: Prompt User for Choice
set "choice="
set /p "choice=Select a model number to launch (1-%count%): "

:: Validate Input
if not defined choice goto MENU
if %choice% lss 1 goto MENU
if %choice% gtr %count% goto MENU

:: Resolve selection variables
set "SELECTED_MODEL_NAME=!model[%choice%]!"
set "SELECTED_MODEL_PATH=!model_path[%choice%]!"
set "MODEL_STEM=%%~nM"

:: Generate paths based on the model file
for %%M in ("!SELECTED_MODEL_PATH!") do (
    set "PRESET_FILE=%%~dpnM.preset"
    set "BASE_STEM=%%~nM"
)

:: ----------------------------------------------------
:: Check for MTP in filename (Case-Insensitive)
:: ----------------------------------------------------
set "MTP_PARAMS="
echo "!SELECTED_MODEL_NAME!" | findstr /I "MTP" >nul 2>&1
if !errorlevel! equ 0 (
    set "MTP_PARAMS=--mmproj-auto --spec-type draft-mtp --spec-draft-n-max 4 --no-mmap --threads 8 --cache-type-k-draft q8_0 --cache-type-v-draft q8_0"
)


:: ----------------------------------------------------
:: Check for a matching mmproj file
:: ----------------------------------------------------
set "MMPROJ_PARAM="
set "base_name=!SELECTED_MODEL_NAME!"

for %%G in (
    -Q2_K -Q2_K_S -Q2_K_L
    -Q3_K_S -Q3_K_M -Q3_K_L -IQ3_XXS -IQ3_S -IQ3_M -IQ3_XS
    -Q4_K_S -Q4_1 -Q4_0 -Q4_K_M -Q4_K_L -Q4_K_XL -IQ4_NL -IQ4_XS
    -Q5_0 -Q5_1 -Q5_K_S -Q5_K_M
    -Q6_K
    -Q8_0 -Q8_1 -Q8_K
    -IQ1_S -IQ1_M -IQ2_S -IQ2_M -IQ2_XXS -IQ2_XS
    -FP16 -BF16 -F16 -F32
) do (
    set "base_name=!base_name:%%G=!"
)
set "base_name=!base_name:.gguf=!"

if exist "%MODEL_DIR%\*!base_name!*mmproj*.gguf" (
    for %%M in ("%MODEL_DIR%\*!base_name!*mmproj*.gguf") do (
        set "MMPROJ_PARAM=--mmproj "%%~fM""
    )
)

:: ----------------------------------------------------
:: Check and Parse Linux-Style .preset File (with # and #old)
:: ----------------------------------------------------
set "CUSTOM_PRESET="
set "skip_rest=false"

if exist "!PRESET_FILE!" (
    set "PRESET_LOADED=true"
    for /f "usebackq delims=" %%L in ("!PRESET_FILE!") do (
        if "!skip_rest!"=="false" (
            set "line=%%L"
            
            :: Strip leading/trailing whitespaces safely
            for /f "tokens=* delims= " %%A in ("!line!") do set "line=%%A"
            
            if defined line (
                :: Check for the hard cutoff token.
                :: Normalize whitespace to handle spaces/tabs between # and old
                set "clean_line=!line!"
                set "clean_line=!clean_line: =!"
                set "clean_line=!clean_line:	=!"
                
                if "!clean_line!"=="#old" (
                    set "skip_rest=true"
                ) else (                    :: Check if the line starts with a regular comment character (#)
                    set "first_char=!line:~0,1!"
                    if not "!first_char!"=="#" (
                        :: Check if the line ends with a backslash
                        if "!line:~-1!"=="\" (
                            set "line=!line:~0,-1!"
                        )
                        :: Append to our preset string
                        set "CUSTOM_PRESET=!CUSTOM_PRESET! !line!"
                    )
                )
            )
        )
    )
) else (
    set "PRESET_LOADED=false"
)


cls
echo ====================================================
echo  Launching llama-server
echo  Model: %SELECTED_MODEL_NAME%
if "!PRESET_LOADED!"=="true" (
    echo  Preset: Loaded Custom Configuration File (.preset)
) else (
    echo  Preset: None found. Using script defaults.
)
if defined MTP_PARAMS (
    echo  Mode: Multi-Token Prediction ^(MTP^) Enabled
)
if defined MMPROJ_PARAM (
    echo  Vision: Multimodal Projector Attached
)
echo ====================================================
echo.

:: :: Execute based on preset existence
:: if "!PRESET_LOADED!"=="true" (
    :: "%SERVER_EXE%" -m "%SELECTED_MODEL_PATH%" !CUSTOM_PRESET! !MTP_PARAMS! !MMPROJ_PARAM!
:: ) else (
    :: "%SERVER_EXE%" -m "%SELECTED_MODEL_PATH%" --seed 0 --host 127.0.0.1 --port 8080 --n-gpu-layers 999 --ctx-size 262144 -np 1 --temp 0.6 --top-p 0.95 --top-k 20 --min-p 0 --keep 4096 --timeout 600 --flash-attn on --presence-penalty 0.8 --repeat-penalty 1.1 --reasoning on --chat-template-kwargs "{\"preserve_thinking\":true}"  --chat-template-file I:\AI\models\LLavacheckpoints\Qwen3.6_chat_template.jinja --cache-type-k q8_0 --cache-type-v q8_0 !MTP_PARAMS! !MMPROJ_PARAM!
:: )

:: Execute based on preset existence
if "!PRESET_LOADED!"=="true" (
    "%SERVER_EXE%" -m "!SELECTED_MODEL_PATH!" !CUSTOM_PRESET!
) else (
    "%SERVER_EXE%" -m "!SELECTED_MODEL_PATH!" --seed 0 --host 127.0.0.1 --port 8080 --n-gpu-layers 999 --ctx-size 262144 -np 1 --temp 0.6 --top-p 0.95 --top-k 20 --min-p 0 --keep 4096 --timeout 600 --flash-attn on --presence-penalty 0.8 --repeat-penalty 1.1 --reasoning on --chat-template-kwargs "{\"preserve_thinking\":true}" --chat-template-file I:\AI\models\LLavacheckpoints\Qwen3.6_chat_template.jinja --cache-type-k q8_0 --cache-type-v q8_0 !MTP_PARAMS! !MMPROJ_PARAM!
)

exit

Fuck Ollama!

Edited September 21 by corz

[argumentum]

On my side I'll do something like this:

Spoiler

Here is the complete AutoIt GUI script. It replaces the console prompt with a window containing a dropdown selector, dynamic configuration preview, and control buttons (Launch, Refresh, and Exit).

 

AutoIt

#pragma compile(Icon, "")
#include <GUIConstantsEx.au3>
#include <ComboConstants.au3>
#include <File.au3>
#include <Array.au3>

; ----------------------------------------------------
; Configuration Pathways
; ----------------------------------------------------
Global Const $SERVER_EXE = "I:\AI\llama.cpp\build\bin\Release\llama-server.exe"
Global Const $MODEL_DIR  = "I:\AI\models\LLavacheckpoints"

; Global Arrays for Model Data
Global $g_aModelName[0]
Global $g_aModelPath[0]

; ----------------------------------------------------
; Create Graphical Interface
; ----------------------------------------------------
Local $hGUI = GUICreate("llama-server Model Launcher", 520, 360, -1, -1)
GUISetFont(9, 400, 0, "Segoe UI")

; Title Header
GUICtrlCreateLabel("llama-server Launcher", 20, 15, 480, 25)
GUICtrlSetFont(-1, 12, 800, 0, "Segoe UI")

; Dropdown Selection
GUICtrlCreateLabel("Select GGUF Model:", 20, 50, 480, 20)
Local $idComboModels = GUICtrlCreateCombo("", 20, 72, 480, 25, $CBS_DROPDOWNLIST)

; Model Status Preview Box
GUICtrlCreateGroup(" Selected Model Configuration ", 20, 115, 480, 175)
Local $idLblPreset  = GUICtrlCreateLabel("Preset File: -", 35, 140, 450, 20)
Local $idLblMtp     = GUICtrlCreateLabel("MTP Mode: -", 35, 168, 450, 20)
Local $idLblVision  = GUICtrlCreateLabel("Vision Projector: -", 35, 196, 450, 20)
Local $idLblStatus  = GUICtrlCreateLabel("Status: Ready", 35, 250, 450, 25)
GUICtrlSetFont($idLblStatus, 9, 700, 0, "Segoe UI")
GUICtrlCreateGroup("", -99, -99, 1, 1)

; Control Buttons
Local $idBtnRefresh = GUICtrlCreateButton("Refresh", 20, 308, 110, 32)
Local $idBtnLaunch  = GUICtrlCreateButton("Launch Server", 250, 308, 140, 32)
GUICtrlSetFont($idBtnLaunch, 9, 700, 0, "Segoe UI")
Local $idBtnExit    = GUICtrlCreateButton("Exit", 400, 308, 100, 32)

; Initial Scan & Render
ScanModels($idComboModels, $idLblStatus)
UpdateModelInfo($idComboModels, $idLblPreset, $idLblMtp, $idLblVision)

GUISetState(@SW_SHOW, $hGUI)

; ----------------------------------------------------
; Main Application Event Loop
; ----------------------------------------------------
While 1
    Local $nMsg = GUIGetMsg()
    Switch $nMsg
        Case $GUI_EVENT_CLOSE, $idBtnExit
            ExitLoop

        Case $idComboModels
            UpdateModelInfo($idComboModels, $idLblPreset, $idLblMtp, $idLblVision)

        Case $idBtnRefresh
            ScanModels($idComboModels, $idLblStatus)
            UpdateModelInfo($idComboModels, $idLblPreset, $idLblMtp, $idLblVision)

        Case $idBtnLaunch
            LaunchServer($idComboModels, $idLblStatus)
    EndSwitch
WEnd

GUIDelete($hGUI)

; ----------------------------------------------------
; Helper Functions
; ----------------------------------------------------

Func ScanModels($idCombo, $idStatus)
    GUICtrlSetData($idCombo, "") ; Clear existing list
    ReDim $g_aModelName[0]
    ReDim $g_aModelPath[0]

    If Not FileExists($MODEL_DIR) Then
        GUICtrlSetData($idStatus, "Status: ERROR - Model directory not found!")
        Return
    EndIf

    Local $hSearch = FileFindFirstFile($MODEL_DIR & "\*.gguf")
    If $hSearch = -1 Then
        GUICtrlSetData($idStatus, "Status: No .gguf models found!")
        Return
    EndIf

    Local $sComboString = ""
    While 1
        Local $sFile = FileFindNextFile($hSearch)
        If @error Then ExitLoop

        ; Exclude vision projectors (mmproj) from the main dropdown
        If Not StringInStr($sFile, "mmproj", 0) Then
            _ArrayAdd($g_aModelName, $sFile)
            _ArrayAdd($g_aModelPath, $MODEL_DIR & "\" & $sFile)
            $sComboString &= $sFile & "|"
        EndIf
    WEnd
    FileClose($hSearch)

    If UBound($g_aModelName) = 0 Then
        GUICtrlSetData($idStatus, "Status: No eligible models found!")
        Return
    EndIf

    ; Remove trailing delimiter
    If StringRight($sComboString, 1) = "|" Then $sComboString = StringTrimRight($sComboString, 1)

    ; Populate Combo Box & Select First Item
    GUICtrlSetData($idCombo, $sComboString, $g_aModelName[0])
    GUICtrlSetData($idStatus, "Status: Found " & UBound($g_aModelName) & " model(s).")
EndFunc

Func GetSelectedIndex($idCombo)
    Local $sSelected = GUICtrlRead($idCombo)
    For $i = 0 To UBound($g_aModelName) - 1
        If $g_aModelName[$i] = $sSelected Then Return $i
    Next
    Return -1
EndFunc

Func UpdateModelInfo($idCombo, $idPreset, $idMtp, $idVision)
    Local $iIndex = GetSelectedIndex($idCombo)
    If $iIndex = -1 Then
        GUICtrlSetData($idPreset, "Preset File: None")
        GUICtrlSetData($idMtp, "MTP Mode: Disabled")
        GUICtrlSetData($idVision, "Vision Projector: None")
        Return
    EndIf

    Local $sSelectedName = $g_aModelName[$iIndex]
    Local $sSelectedPath = $g_aModelPath[$iIndex]

    ; Check for .preset File
    Local $sDrive, $sDir, $sFileName, $sExt
    _PathSplit($sSelectedPath, $sDrive, $sDir, $sFileName, $sExt)
    Local $sPresetFile = $sDrive & $sDir & $sFileName & ".preset"

    If FileExists($sPresetFile) Then
        GUICtrlSetData($idPreset, "Preset File: Custom (" & $sFileName & ".preset)")
    Else
        GUICtrlSetData($idPreset, "Preset File: None (Script defaults will be used)")
    EndIf

    ; Check for MTP
    If StringInStr($sSelectedName, "MTP", 0) Then
        GUICtrlSetData($idMtp, "MTP Mode: Enabled (Multi-Token Prediction)")
    Else
        GUICtrlSetData($idMtp, "MTP Mode: Disabled")
    EndIf

    ; Check for Multimodal Projector
    Local $sMmprojFile = FindMmprojFile($sSelectedName)
    If $sMmprojFile <> "" Then
        GUICtrlSetData($idVision, "Vision Projector: Attached (" & $sMmprojFile & ")")
    Else
        GUICtrlSetData($idVision, "Vision Projector: None Attached")
    EndIf
EndFunc

Func FindMmprojFile($sModelName)
    Local $sBaseName = $sModelName
    Local $aQuantifiers[38] = [ _
        "-Q2_K", "-Q2_K_S", "-Q2_K_L", _
        "-Q3_K_S", "-Q3_K_M", "-Q3_K_L", "-IQ3_XXS", "-IQ3_S", "-IQ3_M", "-IQ3_XS", _
        "-Q4_K_S", "-Q4_1", "-Q4_0", "-Q4_K_M", "-Q4_K_L", "-Q4_K_XL", "-IQ4_NL", "-IQ4_XS", _
        "-Q5_0", "-Q5_1", "-Q5_K_S", "-Q5_K_M", _
        "-Q6_K", _
        "-Q8_0", "-Q8_1", "-Q8_K", _
        "-IQ1_S", "-IQ1_M", "-IQ2_S", "-IQ2_M", "-IQ2_XXS", "-IQ2_XS", _
        "-FP16", "-BF16", "-F16", "-F32" _
    ]

    For $sQuant In $aQuantifiers
        $sBaseName = StringReplace($sBaseName, $sQuant, "", 0, 0)
    Next
    $sBaseName = StringReplace($sBaseName, ".gguf", "", 0, 0)

    Local $hMmprojSearch = FileFindFirstFile($MODEL_DIR & "\*" & $sBaseName & "*mmproj*.gguf")
    If $hMmprojSearch <> -1 Then
        Local $sFile = FileFindNextFile($hMmprojSearch)
        FileClose($hMmprojSearch)
        If Not @error Then Return $sFile
    EndIf
    Return ""
EndFunc

Func ParsePresetFile($sPresetFile)
    Local $sCustomPreset = ""
    If FileExists($sPresetFile) Then
        Local $hFile = FileOpen($sPresetFile, 0)
        If $hFile <> -1 Then
            While 1
                Local $sLine = FileReadLine($hFile)
                If @error Then ExitLoop

                $sLine = StringStripWS($sLine, 3)
                If $sLine <> "" Then
                    Local $sCleanLine = StringReplace(StringReplace($sLine, " ", ""), @TAB, "")
                    If StringLower($sCleanLine) = "#old" Then ExitLoop

                    If StringLeft($sLine, 1) <> "#" Then
                        If StringRight($sLine, 1) = "\" Then $sLine = StringTrimRight($sLine, 1)
                        $sCustomPreset &= " " & $sLine
                    EndIf
                EndIf
            WEnd
            FileClose($hFile)
        EndIf
    EndIf
    Return $sCustomPreset
EndFunc

Func LaunchServer($idCombo, $idStatus)
    Local $iIndex = GetSelectedIndex($idCombo)
    If $iIndex = -1 Then
        MsgBox(16, "Error", "No model selected from the dropdown list.")
        Return
    EndIf

    Local $sSelectedName = $g_aModelName[$iIndex]
    Local $sSelectedPath = $g_aModelPath[$iIndex]

    ; Split path for preset
    Local $sDrive, $sDir, $sFileName, $sExt
    _PathSplit($sSelectedPath, $sDrive, $sDir, $sFileName, $sExt)
    Local $sPresetFile = $sDrive & $sDir & $sFileName & ".preset"

    ; Determine MTP Parameters
    Local $sMtpParams = ""
    If StringInStr($sSelectedName, "MTP", 0) Then
        $sMtpParams = "--mmproj-auto --spec-type draft-mtp --spec-draft-n-max 4 --no-mmap --threads 8 --cache-type-k-draft q8_0 --cache-type-v-draft q8_0"
    EndIf

    ; Determine Multimodal Projector Parameter
    Local $sMmprojFile = FindMmprojFile($sSelectedName)
    Local $sMmprojParam = ""
    If $sMmprojFile <> "" Then
        $sMmprojParam = '--mmproj "' & $MODEL_DIR & "\" & $sMmprojFile & '"'
    EndIf

    ; Parse custom preset file if exists
    Local $sCustomPreset = ParsePresetFile($sPresetFile)

    ; Build Full Executable Command Line
    Local $sFullCmd = ""
    If FileExists($sPresetFile) Then
        $sFullCmd = '"' & $SERVER_EXE & '" -m "' & $sSelectedPath & '"' & $sCustomPreset
    Else
        Local $sDefaultArgs = '--seed 0 --host 127.0.0.1 --port 8080 --n-gpu-layers 999 --ctx-size 262144 -np 1 --temp 0.6 --top-p 0.95 --top-k 20 --min-p 0 --keep 4096 --timeout 600 --flash-attn on --presence-penalty 0.8 --repeat-penalty 1.1 --reasoning on --chat-template-kwargs "{\"preserve_thinking\":true}" --chat-template-file I:\AI\models\LLavacheckpoints\Qwen3.6_chat_template.jinja --cache-type-k q8_0 --cache-type-v q8_0'
        $sFullCmd = '"' & $SERVER_EXE & '" -m "' & $sSelectedPath & '" ' & $sDefaultArgs
        If $sMtpParams <> "" Then $sFullCmd &= " " & $sMtpParams
        If $sMmprojParam <> "" Then $sFullCmd &= " " & $sMmprojParam
    EndIf

    GUICtrlSetData($idStatus, "Status: Launching server for " & $sSelectedName & "...")
    Run($sFullCmd)
EndFunc

Key Enhancements in the GUI Version:

• Dropdown Box (GUICtrlCreateCombo): Automatically populates with all valid models in I:\AI\models\LLavacheckpoints.
• Live Configuration Preview: Selecting any model immediately displays whether a .preset file exists, if MTP draft parameters will be injected, or if a matching vision projector (mmproj) was found.
• Refresh Button: Allows rescanning your folder for newly downloaded models without closing the launcher.
• Non-Blocking Launch: Clicking Launch Server executes llama-server.exe in the background and updates the UI status label without freezing the window.

because I rather GUI everything. ( untested slop for now )

Have to get me "ggerganov/llama.cpp" and become familiar with it.

What is your local hardware setup ? I've got a 5060Ti on this PC and a Strix halo that has more memory, but is slower 😕 

[corz]

1 minute ago, argumentum said:

On my side I'll do something like this:

  Reveal hidden contents

Here is the complete AutoIt GUI script. It replaces the console prompt with a window containing a dropdown selector, dynamic configuration preview, and control buttons (Launch, Refresh, and Exit).

 

AutoIt

 

#pragma compile(Icon, "")
#include <GUIConstantsEx.au3>
#include <ComboConstants.au3>
#include <File.au3>
#include <Array.au3>

; ----------------------------------------------------
; Configuration Pathways
; ----------------------------------------------------
Global Const $SERVER_EXE = "I:\AI\llama.cpp\build\bin\Release\llama-server.exe"
Global Const $MODEL_DIR  = "I:\AI\models\LLavacheckpoints"

; Global Arrays for Model Data
Global $g_aModelName[0]
Global $g_aModelPath[0]

; ----------------------------------------------------
; Create Graphical Interface
; ----------------------------------------------------
Local $hGUI = GUICreate("llama-server Model Launcher", 520, 360, -1, -1)
GUISetFont(9, 400, 0, "Segoe UI")

; Title Header
GUICtrlCreateLabel("llama-server Launcher", 20, 15, 480, 25)
GUICtrlSetFont(-1, 12, 800, 0, "Segoe UI")

; Dropdown Selection
GUICtrlCreateLabel("Select GGUF Model:", 20, 50, 480, 20)
Local $idComboModels = GUICtrlCreateCombo("", 20, 72, 480, 25, $CBS_DROPDOWNLIST)

; Model Status Preview Box
GUICtrlCreateGroup(" Selected Model Configuration ", 20, 115, 480, 175)
Local $idLblPreset  = GUICtrlCreateLabel("Preset File: -", 35, 140, 450, 20)
Local $idLblMtp     = GUICtrlCreateLabel("MTP Mode: -", 35, 168, 450, 20)
Local $idLblVision  = GUICtrlCreateLabel("Vision Projector: -", 35, 196, 450, 20)
Local $idLblStatus  = GUICtrlCreateLabel("Status: Ready", 35, 250, 450, 25)
GUICtrlSetFont($idLblStatus, 9, 700, 0, "Segoe UI")
GUICtrlCreateGroup("", -99, -99, 1, 1)

; Control Buttons
Local $idBtnRefresh = GUICtrlCreateButton("Refresh", 20, 308, 110, 32)
Local $idBtnLaunch  = GUICtrlCreateButton("Launch Server", 250, 308, 140, 32)
GUICtrlSetFont($idBtnLaunch, 9, 700, 0, "Segoe UI")
Local $idBtnExit    = GUICtrlCreateButton("Exit", 400, 308, 100, 32)

; Initial Scan & Render
ScanModels($idComboModels, $idLblStatus)
UpdateModelInfo($idComboModels, $idLblPreset, $idLblMtp, $idLblVision)

GUISetState(@SW_SHOW, $hGUI)

; ----------------------------------------------------
; Main Application Event Loop
; ----------------------------------------------------
While 1
    Local $nMsg = GUIGetMsg()
    Switch $nMsg
        Case $GUI_EVENT_CLOSE, $idBtnExit
            ExitLoop

        Case $idComboModels
            UpdateModelInfo($idComboModels, $idLblPreset, $idLblMtp, $idLblVision)

        Case $idBtnRefresh
            ScanModels($idComboModels, $idLblStatus)
            UpdateModelInfo($idComboModels, $idLblPreset, $idLblMtp, $idLblVision)

        Case $idBtnLaunch
            LaunchServer($idComboModels, $idLblStatus)
    EndSwitch
WEnd

GUIDelete($hGUI)

; ----------------------------------------------------
; Helper Functions
; ----------------------------------------------------

Func ScanModels($idCombo, $idStatus)
    GUICtrlSetData($idCombo, "") ; Clear existing list
    ReDim $g_aModelName[0]
    ReDim $g_aModelPath[0]

    If Not FileExists($MODEL_DIR) Then
        GUICtrlSetData($idStatus, "Status: ERROR - Model directory not found!")
        Return
    EndIf

    Local $hSearch = FileFindFirstFile($MODEL_DIR & "\*.gguf")
    If $hSearch = -1 Then
        GUICtrlSetData($idStatus, "Status: No .gguf models found!")
        Return
    EndIf

    Local $sComboString = ""
    While 1
        Local $sFile = FileFindNextFile($hSearch)
        If @error Then ExitLoop

        ; Exclude vision projectors (mmproj) from the main dropdown
        If Not StringInStr($sFile, "mmproj", 0) Then
            _ArrayAdd($g_aModelName, $sFile)
            _ArrayAdd($g_aModelPath, $MODEL_DIR & "\" & $sFile)
            $sComboString &= $sFile & "|"
        EndIf
    WEnd
    FileClose($hSearch)

    If UBound($g_aModelName) = 0 Then
        GUICtrlSetData($idStatus, "Status: No eligible models found!")
        Return
    EndIf

    ; Remove trailing delimiter
    If StringRight($sComboString, 1) = "|" Then $sComboString = StringTrimRight($sComboString, 1)

    ; Populate Combo Box & Select First Item
    GUICtrlSetData($idCombo, $sComboString, $g_aModelName[0])
    GUICtrlSetData($idStatus, "Status: Found " & UBound($g_aModelName) & " model(s).")
EndFunc

Func GetSelectedIndex($idCombo)
    Local $sSelected = GUICtrlRead($idCombo)
    For $i = 0 To UBound($g_aModelName) - 1
        If $g_aModelName[$i] = $sSelected Then Return $i
    Next
    Return -1
EndFunc

Func UpdateModelInfo($idCombo, $idPreset, $idMtp, $idVision)
    Local $iIndex = GetSelectedIndex($idCombo)
    If $iIndex = -1 Then
        GUICtrlSetData($idPreset, "Preset File: None")
        GUICtrlSetData($idMtp, "MTP Mode: Disabled")
        GUICtrlSetData($idVision, "Vision Projector: None")
        Return
    EndIf

    Local $sSelectedName = $g_aModelName[$iIndex]
    Local $sSelectedPath = $g_aModelPath[$iIndex]

    ; Check for .preset File
    Local $sDrive, $sDir, $sFileName, $sExt
    _PathSplit($sSelectedPath, $sDrive, $sDir, $sFileName, $sExt)
    Local $sPresetFile = $sDrive & $sDir & $sFileName & ".preset"

    If FileExists($sPresetFile) Then
        GUICtrlSetData($idPreset, "Preset File: Custom (" & $sFileName & ".preset)")
    Else
        GUICtrlSetData($idPreset, "Preset File: None (Script defaults will be used)")
    EndIf

    ; Check for MTP
    If StringInStr($sSelectedName, "MTP", 0) Then
        GUICtrlSetData($idMtp, "MTP Mode: Enabled (Multi-Token Prediction)")
    Else
        GUICtrlSetData($idMtp, "MTP Mode: Disabled")
    EndIf

    ; Check for Multimodal Projector
    Local $sMmprojFile = FindMmprojFile($sSelectedName)
    If $sMmprojFile <> "" Then
        GUICtrlSetData($idVision, "Vision Projector: Attached (" & $sMmprojFile & ")")
    Else
        GUICtrlSetData($idVision, "Vision Projector: None Attached")
    EndIf
EndFunc

Func FindMmprojFile($sModelName)
    Local $sBaseName = $sModelName
    Local $aQuantifiers[38] = [ _
        "-Q2_K", "-Q2_K_S", "-Q2_K_L", _
        "-Q3_K_S", "-Q3_K_M", "-Q3_K_L", "-IQ3_XXS", "-IQ3_S", "-IQ3_M", "-IQ3_XS", _
        "-Q4_K_S", "-Q4_1", "-Q4_0", "-Q4_K_M", "-Q4_K_L", "-Q4_K_XL", "-IQ4_NL", "-IQ4_XS", _
        "-Q5_0", "-Q5_1", "-Q5_K_S", "-Q5_K_M", _
        "-Q6_K", _
        "-Q8_0", "-Q8_1", "-Q8_K", _
        "-IQ1_S", "-IQ1_M", "-IQ2_S", "-IQ2_M", "-IQ2_XXS", "-IQ2_XS", _
        "-FP16", "-BF16", "-F16", "-F32" _
    ]

    For $sQuant In $aQuantifiers
        $sBaseName = StringReplace($sBaseName, $sQuant, "", 0, 0)
    Next
    $sBaseName = StringReplace($sBaseName, ".gguf", "", 0, 0)

    Local $hMmprojSearch = FileFindFirstFile($MODEL_DIR & "\*" & $sBaseName & "*mmproj*.gguf")
    If $hMmprojSearch <> -1 Then
        Local $sFile = FileFindNextFile($hMmprojSearch)
        FileClose($hMmprojSearch)
        If Not @error Then Return $sFile
    EndIf
    Return ""
EndFunc

Func ParsePresetFile($sPresetFile)
    Local $sCustomPreset = ""
    If FileExists($sPresetFile) Then
        Local $hFile = FileOpen($sPresetFile, 0)
        If $hFile <> -1 Then
            While 1
                Local $sLine = FileReadLine($hFile)
                If @error Then ExitLoop

                $sLine = StringStripWS($sLine, 3)
                If $sLine <> "" Then
                    Local $sCleanLine = StringReplace(StringReplace($sLine, " ", ""), @TAB, "")
                    If StringLower($sCleanLine) = "#old" Then ExitLoop

                    If StringLeft($sLine, 1) <> "#" Then
                        If StringRight($sLine, 1) = "\" Then $sLine = StringTrimRight($sLine, 1)
                        $sCustomPreset &= " " & $sLine
                    EndIf
                EndIf
            WEnd
            FileClose($hFile)
        EndIf
    EndIf
    Return $sCustomPreset
EndFunc

Func LaunchServer($idCombo, $idStatus)
    Local $iIndex = GetSelectedIndex($idCombo)
    If $iIndex = -1 Then
        MsgBox(16, "Error", "No model selected from the dropdown list.")
        Return
    EndIf

    Local $sSelectedName = $g_aModelName[$iIndex]
    Local $sSelectedPath = $g_aModelPath[$iIndex]

    ; Split path for preset
    Local $sDrive, $sDir, $sFileName, $sExt
    _PathSplit($sSelectedPath, $sDrive, $sDir, $sFileName, $sExt)
    Local $sPresetFile = $sDrive & $sDir & $sFileName & ".preset"

    ; Determine MTP Parameters
    Local $sMtpParams = ""
    If StringInStr($sSelectedName, "MTP", 0) Then
        $sMtpParams = "--mmproj-auto --spec-type draft-mtp --spec-draft-n-max 4 --no-mmap --threads 8 --cache-type-k-draft q8_0 --cache-type-v-draft q8_0"
    EndIf

    ; Determine Multimodal Projector Parameter
    Local $sMmprojFile = FindMmprojFile($sSelectedName)
    Local $sMmprojParam = ""
    If $sMmprojFile <> "" Then
        $sMmprojParam = '--mmproj "' & $MODEL_DIR & "\" & $sMmprojFile & '"'
    EndIf

    ; Parse custom preset file if exists
    Local $sCustomPreset = ParsePresetFile($sPresetFile)

    ; Build Full Executable Command Line
    Local $sFullCmd = ""
    If FileExists($sPresetFile) Then
        $sFullCmd = '"' & $SERVER_EXE & '" -m "' & $sSelectedPath & '"' & $sCustomPreset
    Else
        Local $sDefaultArgs = '--seed 0 --host 127.0.0.1 --port 8080 --n-gpu-layers 999 --ctx-size 262144 -np 1 --temp 0.6 --top-p 0.95 --top-k 20 --min-p 0 --keep 4096 --timeout 600 --flash-attn on --presence-penalty 0.8 --repeat-penalty 1.1 --reasoning on --chat-template-kwargs "{\"preserve_thinking\":true}" --chat-template-file I:\AI\models\LLavacheckpoints\Qwen3.6_chat_template.jinja --cache-type-k q8_0 --cache-type-v q8_0'
        $sFullCmd = '"' & $SERVER_EXE & '" -m "' & $sSelectedPath & '" ' & $sDefaultArgs
        If $sMtpParams <> "" Then $sFullCmd &= " " & $sMtpParams
        If $sMmprojParam <> "" Then $sFullCmd &= " " & $sMmprojParam
    EndIf

    GUICtrlSetData($idStatus, "Status: Launching server for " & $sSelectedName & "...")
    Run($sFullCmd)
EndFunc

Key Enhancements in the GUI Version:

• Dropdown Box (GUICtrlCreateCombo): Automatically populates with all valid models in I:\AI\models\LLavacheckpoints.
• Live Configuration Preview: Selecting any model immediately displays whether a .preset file exists, if MTP draft parameters will be injected, or if a matching vision projector (mmproj) was found.
• Refresh Button: Allows rescanning your folder for newly downloaded models without closing the launcher.
• Non-Blocking Launch: Clicking Launch Server executes llama-server.exe in the background and updates the UI status label without freezing the window.

because I rather GUI everything. ( untested slop for now )

Have to get me "ggerganov/llama.cpp" and become familiar with it.

What is your local hardware setup ? I've got a 5060Ti on this PC and a Strix halo that has more memory, but is slower 😕 

RTX 3060 12GB. 

I GGUF everything!

[corz]

Just now, corz said:

RTX 3060 12GB. 

I GGUF everything!

Also, llama.cpp is easier than you probably think. Rather than explain, see here.

[mLipok]

6 hours ago, corz said:

Also, that shit needs to get sent with EVERY prompt.

With ChatGPT codex this is taken automaticaly.
With chatty I put :

Quote

Prepare the project in a GitHub-compatible format as a zip archive.
Name the zip file using the following format:
YYYY-MM-DD__HHMM_vXXX__[projectname].zip
Number the versions (XXX) sequentially, starting from the very beginning of the project.
Include the version date and version number within the project in this format:
$_PROJECT_VERSION = "YYYY-MM-DD__HHMM_vXXX__[projectname].zip"

Do not make immediate corrections; instead, add them to the next iteration and execute the iteration only when I request it.
After discussing a task and adding it to the next iteration, provide the current list of items for that iteration.
Add all agreements—including this one—to AGENTS.md.

And then I start to co-work with AI

 

 

Edited September 21 by mLipok

[TheDcoder]

I'm curious, how often does AI make mistakes when you use it to write code? In my limited experiments, it tends to make a lot of unnecessary assumptions and write non-functional code that I often have to debug, I imagine the experience for y'all is different because I can't imagine using it in the current state.

One thing it does an acceptable job at is writing interface as mentioned by corz, and I don't mind if the code is not correct because the UI isn't a critical piece of infrastructure and it's easy enough to test.

[mLipok]

I have big 2 projects writen entirly with AI 
one with C# one with AutoIt
Both really big working on it about 1 month
Have many integrations

btw. recently 2 udfs posted by me here on the forum are also created with AI

and next will come soon

 

[mLipok]

59 minutes ago, TheDcoder said:

I'm curious, how often does AI make mistakes when you use it to write code?

If you calculate it based on the time saved,
I’d say it’s equivalent to a month’s worth of work as a programmer without AI.

Unless you’re writing bad prompts—in which case, it’s several times a day (this time in terms of actual clock time).

[TheDcoder]

2 minutes ago, mLipok said:

Both really big working on it about 1 month

How was the development process, was it mostly tell it what to do (hands-off) or was it more like working with a forgetful co-worker who needs a lot of hand-holding? Or to be more specific, where did it fall on the scale between those two experiences

[mLipok]

1 hour ago, TheDcoder said:

where did it fall on the scale between those two experiences

I feel this way:

 

Quote

I am a Product Owner.
I have meetings where I discuss development directions and bugs with the lead AI person (Software Architect).
As a rule, I don't touch the code.
I have touched it a few times to fix minor issues—like specifying a `#forceref`—before moving on to verify the delivered fix.

 

[mLipok]

Hello every body in this conversation.
I encourage you to read more here:

 

Edited Saturday at 08:56 PM by mLipok

[corz]

On 9/21/2026 at 12:20 PM, TheDcoder said:

I'm curious, how often does AI make mistakes when you use it to write code? In my limited experiments, it tends to make a lot of unnecessary assumptions and write non-functional code that I often have to debug, I imagine the experience for y'all is different because I can't imagine using it in the current state.

One thing it does an acceptable job at is writing interface as mentioned by corz, and I don't mind if the code is not correct because the UI isn't a critical piece of infrastructure and it's easy enough to test.

It depends on the code and the model.

Some models hit walls with certain kids of tasks. Muse Spark is my "Junior engineer", and gets all the grunt work to do as I couldn't burn through its allowance if I tried. On high it's on-par with a 2nd year apprentice, and so long as you know your codebase and can spot obvious stupidity, it will produce decent code. Sometimes excellent code (on High, always).

There's a big difference (in tokens) between building a thing from scratch and fixing a thing that's got a few bugs. When complete, hand the work off to Sol 6 medium (et al) and say, "Check out my work on the Gobungle() function. I think I fucked-up somewhere. Can you see it?"

Or Astra Max/Ultimate (the purple one!) if it's really fucked up. Using better models is also a good way to PLAN a bigger change. They will sometimes spot architectural issues that lesser models miss altogether. I had a brief fling with GLM 5.3 and while was impressed by it's grasp of large codebases and big "concepts", it codes like a baby and hasn't been used since. Find you favourite by working with it on some sample projects, some models/harnesses click with you and your way of working some don't.

So you can have Sol/Astra come up with a rock-solid plan, then hand that exact plan to the cheap model (with all points clear so it doesn't fuck up), then hand the "finished" result (your project files as they are now) back to a better model to "double-check". I usually say "double-check your work", even when the work was done by another model as it makes them look harder for bugs and work harder to provide better solutions if they do find bugs, watching them think reveals stuff like: I can't believe I wrote that code! I missed two obvious logic holes!

Muse often surprises me in both good and bad ways. I'll have it left on Build by mistake and simply mention a thing and before I know it, it's already refactored the entire codebase to incorporate the new "idea"! Erm... OKAY! And the code is rock-solid. The next time it misses obvious shit. You need to know your code to make the most of it. Letting it loose without reviewing the logic or understanding the changes is not yet wise. But soon! 

Also Muse HATES to be in plan mode (OpenCode has 2 basic agents Plan & Build, plan disallows file changes) and stops at nothing to just go implement whatever you desire, like a little puppy with it's waggy tail, "Done!", it says, sometimes in amazing speeds. I hit F8 in Notepad++ and a few seconds later I can go, "FFS! NO! EQUAL SPACING YOU FUCKWIT!!!!" [thinking: user is getting frustrated at my failure to measure window accurately]. Gotta love vibe coding! Building this shit on LLM's was the right move!

Muse is basically free at this time and it's like having a free employee. Dunno how long it will last. You can open multiple agents simultaneously and have it working on multiple projects and you would still need to have some sort of time-travel device to burn through your token allowance! I've been at it for maybe eight hours today and burned like $1.50, and that's of my "virtual" $60 allowance, which costs £7.50 / month! Fkn nuts!

There are a couple coding tasks that sent it into the equivalent of robot-hitting-head-against-wall, and after a few turns you realise it's just got a blockage there. And like EM Dashes, there's certain coding structures and data structures that it prefers over others, no matter how lazy or ugly they are. So you need to be on your toes! But it's a lot of fun. Way more than it will be soon when it just "does it" for you, like Sol/Astra can do mostly right now, except for 1000x the cost.

Horses for courses. Sol Medium is my "Senior Engineer" (or one of its brothers), and if shit goes sideways he'll definitely see the err of "my" stupid ways. OpenAi models also love to be the hero, so if you say, "OMFG! I let some cheap AI agent loose on my code and totally fucked up the ini parsing!", or whatever is broken, it will work extra hard to save you from evils of that "lesser" intelligence.

Pick the right model for the right task, is my best advice.

And make your prompt EXCELLENT.

I once asked an AI, after a particularly productive session, something along the lines of, "I realise you have no emotions, but I humanly feel the need to say thank you. What's the best way to thank an AI?", and it's response has guided every prompt I've written since. It said, roughly, "By providing accurate prompts which reduce my need to infer, precision is important and efficient", and so on. And I got it! I can see the robot banging its head against the wall as its AI brain traverses the layers of the matrix calculations going NO!!!!!!! NO!!!!!!!!!!! NO!!!!!!!!!!! Poor thing!

Great prompting is crucial with cheaper models. If you can create a prompt that leaves no holes, guards all rails, makes the goals precise and concrete, guaranteed, loads of free models available right now can write you solid AutoIt code.

 

[edit]Muse Spark 1.3, time of writing model has a HUGE, I mean MASSIVE context window, 1 million tokens. DO NOT let that fool you. If the wheel goes past 50%, you need to wrap up and start a new thread. You can ask it to summarise for you and hand that off to a fresh session. Past three quarters (750,000) tokens it is starting to adopt whole different ways of speaking and coding, reverting to machine-speak and wonky archaic structures, and fuck your AGENTS.md! If you, thinking, hey, I'm smart, I'll just /compact your context history, well, God Help You! Not saying you might have fun, but not coding past a fresh 50% is has become my number one Muse rule.[/edit]

Edited Sunday at 11:56 PM by corz

[corz]

5 minutes ago, corz said:

"Senior Engineer".

 

I should add.

You don't actually get the senior engineer to fix the issues! His rates are too high! You hand his verdict verbatim back to the junior engineer who goes, "You are right! I should have seen that! New plan! ..."

Then when it gets it you switch to build (Ctrl+., TAB in the CLI version) and it does it. You switch back to senior engineer and say, "What about that?", and if there's still issues, it will tell you. 

NEVER setup two AI in an automatic adversarial situation and expect great code, not happening. But with you in the middle you can get stuff good-as-can-be for minimal tokens.

[argumentum]

I found that, using Gemini AI free, if it messed up and I open a new tab, telling it that "I coded this but sucks 😭", it'll tell me to not cry, here are the mistakes and "this is the excellent, perfect, am the best, running code" ( because that thing can have an ego  ) but, if I claim that it messed up, nope, it's gonna defend the faulty code 🤷‍♂️

So if you take the blame, it'll help you, but if you say "you messed up", it'll go into blame mode and freeze/traumatize.

And it has its days. There are days that is clear minded, and days that you'd look for a way to kick that thing in the RAM ( it has no balls ).

For what it told me regarding the free plans and the paid ones, is the same KV allotment, same everything in regards to thinking patterns. So is not better if paid, just more tokens to continue.