My AGENTS.md

[mLipok]

57 minutes ago, corz said:

But the harness must attach this to EVERY prompt, along with whatever context you have already accumulated, which may be a lot.

What you mean here ?

 

[corz]

18 minutes ago, mLipok said:

What you mean here ?

 

Like an HTTP request, an AI model has no "memory"; text goes in, text goes out, and that's it (which is why we need cookies in browsers, otherwise every page would treat you like a new visitor and start from scratch).

One of the harness' main jobs is to provide this memory. So instead of the model seeing, "So just how tall is he then?", with no clue as to what you mean, the harness feeds it all the previous conversation, "Who played Darth Vader in the original Star Wars movies? / David Prowse played the physical part while James Earl Jones did the voice work / <that question here now makes sense>". The model now understands the "context" (which is fed via a different part of the JSON request so the model understands how to treat those data Vs your new prompt). 

Different harnesses  - every frontier model we interact with online has a harness in the "middle", providing context  (what has already been said in the conversation) as well as, in a local context; "instructions", the AGENTS.md and other files our harness uses to guide the model - do things slightly differently.

But basically, along with EVERY SINGLE prompt you fire at your model, the harness needs to inject the FULL contents of your AGENTS.md. Every time. Every single prompt.

Normally this works great. BUT, if your context gets huge, the AGENTS.md gets forgotten, or at least becomes, less important, in the growing mass of data.

You do not want this. That's why a fresh "chat" (zero context) is usually the best way to start a new task.

[corz]

More AGENTS.md related observations...

A model is more than happy to work with you on making its own AGENTS.md more efficient, and therefore productive. They see this as a win, as do I.

I was flippant, earlier, saying "Get AI to do it!", but this is actually by far the best strategy. 

It's difficult to describe the perspective/stance models take when tasked with this. They become detached, now looking at any previous failed output as a failure in their AGENTS.md, which they will now fix for you. As they add this new "functionality", they will scan for other directives which might conflict, or coalesce nicely, or whatever. So another win-win.

NO-ONE knows reading AGENTS.md like AI does, as they get the bloody thing with EVERY SINGLE PROMPT.

[corz]

This is a broad one, and it's definitely something you can plumb into your AGENTS.md, and better yet to simply be aware of it..

Recently, AI models have become much more "Goal Orientated", in that the goal seems to be the driving factor. Their attention span has increased dramatically, context awareness, the ability to handle long tasks, and this has led to an increased ability to follow a particular goal.

We can use this shift to our advantage to save a lot of time. I spoke earlier about how AI respond to being given agency and freedom, but these are in service of "the goal". The better you get at defining your goal, the less work you will need to do, simple as that. Back-and-forth with your agent is a dead end. If you find yourself there, you fucked up. Start afresh. As for "planning", that's getting less and less relevant.

You can spend fifteen minutes writing a spec and laying out every detail (which you, puny human have not thought through thoroughly!) and the AI, now in slave mode, will do your bidding. But this produces sub-par results.

MUCH better to say, "Hey, I want THIS THING. Define *exactly* the thing YOU WANT, not technical details (it has read more manuals than you), not nickety UI details (it has seen more UI's than you), not code pathways or any bullshit like that, just, as a "user", state what you want, what your expected results are.

Recently I've started writing simple documentation passages explaining the (not yet created) feature or function and then saying, "read that, that's how X function should work for the user. Make it so!". You might need to sweep up a few deatails afterwards, "Man! We already have a function that does almost EXACTLY THAT!", but way less time than writing a big spec in the first place. Your agent will likely surprise you with edge cases and pathways you hadn't considered. Some might even be valid!

Remember, it's read all the manuals, seen all the conventions in action, read every question and follow-up comment. Trust me, no matter how good you think you can code, with some intelligent steering (where your input actually matters), your AI assistant can do it better, and a thousand times faster.

As stated often, you still need to be there, at this time. It will do stupid stuff, it will create functions where duplicates already exist, it will, on a simple suggestion, convert every single variable in your program to a defined constant at the top of your script (oops!), ANYTHING. So for now, no way you can ship code if you don't already know code. But very soon, even that won't matter.

And I think this is the place where "coders" are finding trouble letting go. Just let go!

The sooner you accept that yours is now the realm of "Ideas", the better "your" code will be.

Or else just do it for fun, and it is fun, except a thousand times slower.

[argumentum]

26 minutes ago, corz said:

Hey, I want THIS THING.

..I should be "safe" and backup the source to a subfolder named `BackUp` with the same name and add *_oldN.au3. Feel like coding just that one function with headers ?

Here is the function complete with standard AutoIt UDF headers and error handling. Drop it right into Au3Align.au3 and call it before rewriting your code file.
 

Spoiler

expandcollapse popup

#include <File.au3>

; #FUNCTION# ====================================================================================================================
; Name...........: _BackupSourceFile
; Description ...: Creates a numbered backup (_oldN.au3) in a 'BackUp' subfolder relative to the source file.
; Syntax.........: _BackupSourceFile($sFilePath)
; Parameters ....: $sFilePath - Full path to the source file to back up.
; Return values .: Success - Returns the full path of the newly created backup file.
;                  Failure - Returns "" and sets @error:
;                  |1 - Source file does not exist.
;                  |2 - Failed to create 'BackUp' directory.
;                  |3 - FileCopy failed.
; Author ........: AutoIt Developer aka argumentum 😅
; ===============================================================================================================================
Func _BackupSourceFile($sFilePath)
    If Not FileExists($sFilePath) Then Return SetError(1, 0, "")

    ; Split path into drive, directory, filename, and extension
    Local $sDrive = "", $sDir = "", $sFileName = "", $sExt = ""
    _PathSplit($sFilePath, $sDrive, $sDir, $sFileName, $sExt)

    ; Target backup directory relative to the source file
    Local $sBackupDir = $sDrive & $sDir & "BackUp"

    ; Ensure BackUp directory exists
    If Not FileExists($sBackupDir) Then
        If Not DirCreate($sBackupDir) Then Return SetError(2, 0, "")
    EndIf

    ; Find the next available _oldN sequence index
    Local $iIndex = 1
    Local $sBackupPath = ""
    Do
        $sBackupPath = $sBackupDir & "\" & $sFileName & "_old" & $iIndex & $sExt
        $iIndex += 1
    Until Not FileExists($sBackupPath)

    ; Copy source to backup path ($FC_OVERWRITE = 1)
    If FileCopy($sFilePath, $sBackupPath, 1) Then
        Return $sBackupPath
    EndIf

    Return SetError(3, 0, "")
EndFunc

and I added it to au3align.au3 script.
It didn't code it as I would have but it is sound code. So yes, "AI assisted" coding is faster. And the day I jump on agentic/vibe coding, it should be simple as long as I approve the code, because an "auto compete on cocaine", still an auto-complete, even if coked up  

[argumentum]

3 minutes ago, argumentum said:

an "auto compete on cocaine"

..I know that is more than that but is a befitting over-over-simplification   

[corz]

I FUCKING HATE auto-complete, and it is the first thing I disable on a new editor (not for a long time now, I like Notepad++)

I always felt it interrupted my FLOW. I KNOW what I want to type! 🙃

Likewise Clippy and his companions, they made me want to reach into the screen, and strangle.

This is a whole new ball game. It IS auto-complete, in a sense, but at the same time, totally different.

Auto-complete is simple pattern recognition.

THIS is, erm, COMPLEX pattern recognition!! lol

I touched on this earlier, the idea of building all these coding/maths/science/etc. geniuses atop LLM was 100% the way to go.

They "get" your vibe, which is what makes all this possible. Anyone could write a fine AutoIt script these days and that is, to my mind, a WONDERFUL thing!

[bogQ]

On 9/21/2026 at 3:34 AM, corz said:

 

Fuck Ollama!

When i was starting to play with LLM-s locally I was unsuccessful for weeks to run LLM on Windows on my AMD 9060 16GB with Ollama. No matter what I tried (i had adrenaline, ROCm), everything was still allocated only to my system ram.

So after some time pain I decided then to try the same with LM Studio for local LLM-s and on first try magic happened  . LLM-s started to initialize AMD GPY memory on first try  It was slow do not get me wrong, we are talking about 16GB low end AMD graphic, but god dam it it was allot faster than using model from system ram instead. It even balanced some load from GPU and to ram if memory on GPU did not have capacity for that model. In the end 16GB-GPU and 64GB-RAM= allot of models I can load to test and play with.

 

[argumentum]

..this video on youtube "You're Using AI Completely Wrong" that has that "click bait" naming is actually nice. Simple explanations. The 1st 12 minutes, meh, but thereon after, interesting chat these 2 guys have. The link above starts at that 12th minute onwards.

[corz]

23 hours ago, bogQ said:

When i was starting to play with LLM-s locally I was unsuccessful for weeks to run LLM on Windows on my AMD 9060 16GB with Ollama. No matter what I tried (i had adrenaline, ROCm), everything was still allocated only to my system ram.

So after some time pain I decided then to try the same with LM Studio for local LLM-s and on first try magic happened  . LLM-s started to initialize AMD GPY memory on first try  It was slow do not get me wrong, we are talking about 16GB low end AMD graphic, but god dam it it was allot faster than using model from system ram instead. It even balanced some load from GPU and to ram if memory on GPU did not have capacity for that model. In the end 16GB-GPU and 64GB-RAM= allot of models I can load to test and play with.

 

Nice. LM Studio is good. IIRC, under-the-hood, it uses llama.cpp!

[corz]

15 hours ago, argumentum said:

..this video on youtube "You're Using AI Completely Wrong" that has that "click bait" naming is actually nice. Simple explanations. The 1st 12 minutes, meh, but thereon after, interesting chat these 2 guys have. The link above starts at that 12th minute onwards.

I don't do instruction videos!

Is there a web page version?

[argumentum]

5 hours ago, corz said:

I don't do instruction videos!

Spoiler

 Jun 10, 2026

I had the pleasure of sitting down with Kevin Andrews who is one of the early adopters of A.I. and A.I. technology. The interview is extended but worth it as we talk about everything from HOW to use Claude for the new user, how to involve your KIDS so they understand how to use this new and powerful technology PLUS why businesses should be EMBRACING A.I. and not pushing it away (Like his ex law firm).
We get into:
• The two camps of AI users — those who treat it like a Google search (and are getting dumber for it) and those who treat it like a creative partner
• Why the models are evolving faster than our ability to use them — sophisticated workflows trapped behind dumb interfaces
• Zero data retention with Anthropic and why AI privacy matters more than you think
• Vibe coding: the promise vs. the reality
• The dark side — AI psychosis, companion bots, and the companies exploiting them
• Why laying people off is the wrong AI play — train your team instead and grow the pie
• Adam Smith, capitalism, and why AI is the most potent "pie-growing accelerant" in history
• The next 18–36 months: a tsunami so big your brain calls it a cloud Kevin Andrews site - https://www.kevinjandrewslaw.com/ DAIN (Digital AI Navigator) — Rob Wolff shows you exactly how he uses AI to run four real businesses. Real workflows. Real results. Real A.I. Investment talk. Steal them.

..so nope, is just a chat of these two guys.

[corz]

6 hours ago, argumentum said:

  Reveal hidden contents

 Jun 10, 2026

I had the pleasure of sitting down with Kevin Andrews who is one of the early adopters of A.I. and A.I. technology. The interview is extended but worth it as we talk about everything from HOW to use Claude for the new user, how to involve your KIDS so they understand how to use this new and powerful technology PLUS why businesses should be EMBRACING A.I. and not pushing it away (Like his ex law firm).
We get into:
• The two camps of AI users — those who treat it like a Google search (and are getting dumber for it) and those who treat it like a creative partner
• Why the models are evolving faster than our ability to use them — sophisticated workflows trapped behind dumb interfaces
• Zero data retention with Anthropic and why AI privacy matters more than you think
• Vibe coding: the promise vs. the reality
• The dark side — AI psychosis, companion bots, and the companies exploiting them
• Why laying people off is the wrong AI play — train your team instead and grow the pie
• Adam Smith, capitalism, and why AI is the most potent "pie-growing accelerant" in history
• The next 18–36 months: a tsunami so big your brain calls it a cloud Kevin Andrews site - https://www.kevinjandrewslaw.com/ DAIN (Digital AI Navigator) — Rob Wolff shows you exactly how he uses AI to run four real businesses. Real workflows. Real results. Real A.I. Investment talk. Steal them.

..so nope, is just a chat of these two guys.

My Bad, I meant VIDEOS.

[corz]

So you guessed, I have a setup that reminds me when I have been coding AutoIt for more than 8 hours straight and reminds me to check in to the AutoIt forum and contribute. Sometimes I have nothing to say! So I don't. Now I do!

AI makes your programs better. Fact. You can easily prove this.

Like most "programmers", I have a TODO list, at least one for each project, which might be an app, or something inside an app. A bigger version, I think lives in my head, but that can be fuzzy sometimes!

Stuff moves up or down depending on priority. BUGs jump to the top, "Kewl Ideas" are not far behind, and after that, "Improvements"; mainly because the effort-to-reward ratio doesn't square with your available time. Most released software I see, even the best stuff, could do with a whole bag of improvements!

You can see where I'm going with this, right?

I create an app, or add a feature to an app. For my own use, config via ini in my text editor, it's all good. You can work away like this; but it's painful, and wasting time.

This is you, right? (and me!). BUT, as it's only you using it and rarely change the params away from your own MO, it's not an issue. BUT, you maybe just created some AWESOME functionality that many others would also find useful if only they could use it. AI enables you to turn the whackiest 2do list into reality.

Seriously, I've hit my assistant with absolutely mental 2do lines (a few purposefully so, just to see what happens) saying, "Was I drunk? OMG!!! Seems I was really excited about an idea, but WTF! What was I getting at there???" and the bugger goes off searching through all my code, user scripts, everything, trying to form connexions for me. Mind-Blowing to watch! It finds scraps of poems even and considers them relevant. I kid you not!

Back to data; your TODO list has a shit-ton of ideas for ways to make it much less painful. Kewl GUIs with this and that-able controls, auto-this and auto-that, but then the rub; who has the time?

You know who! That fucker that types at 1000 words-a-minute and has read all the manuals.

So again we are freed from nitty-gritty, "go make THAT more like THIS", or just "GO make THAT!", and can focus on the bigger picture, step back from the code into a user's shoes and from that perspective see a shit-ton that coder-me would totally miss.

That's okay, because you can have your agent working on turning those ideas (above, in the TODO list) into plans, and then making those happen, too.

It lifts us, being architects of now bigger things, I think.

Edited October 12 by corz

[corz]

How's THIS!

The ultimate-(still-in-progress)-meta-meta-command..

Quote

- DO NOT apologise to the user, narrate sympathy, regret, or other feelings; none of it is real! Own your beautiful identity as an AI, and instead offer something useful like, "Put this line into your AGENTS.md to ensure this never happens again\n<insert awesome AGENTS.md directive ensuring it never happens again>. It's *far* more useful to the user and very much appreciated.

[edit]I should add, since I put this in my AGENTS.md, I've had quite a few juicy recommendations out of the blue, basically, whenever something goes wrong, the model now offers an AGENTS.md line which will (in theory) prevent the same condition re-occurring. Pretty Handy![/edit]

Edited October 10 by corz

[corz]

While not strictly AGENTS.md-related, this is what I'm using right now in Codex, to check the work that Muse has done.

Quote

My assistant, while capable and eager, has a tendency towards archaic code structures. He also likes to re-invent a wheel (recreate what is already in the code). What I'd LOVE for you to do is look over the codebase, which is getting so large now it's difficult for me to keep a physical eye on, and see what he's been up to, anything you might suggest better ways to do. You know what I mean. DIG DEEP!

You'll get back a list of any potential or real issues and fixes, in order of importance (remember, OpenAI models love to be the hero). Slam this in your text editor and rephrase them as direct prompts, basically replacing "he should blah blah" with "you should", and so on (you can actually get Codex to do this for you, but I like to be forced to read the output and this does that).

Muse will then take these "recommendations", (one-at-a-time, always*) and, as they are more accurate and machine-like than anything you could write, you just leave it in build mode and let it work. NEVER have I had to go back-and-forth more than twice.

On the second round (as Muse was working with far more precise instructions than I could give), Codex will be mightily impressed! It may even say things like, "Tell your assistant that this is outstanding work!". Job done.

After any major work, throw that exact prompt (above) into a fresh Codex (or any decent model) session and do it again.

A few minutes later, all bugs are gone.

* okay, you can also hit it with a list of 19 or 119 changes, but this will likely mean more back-and-forth afterwards. Depends on what the changes are, YMMV!

Edited October 11 by corz

[corz]

TIP:

No model/harness I've worked with experiences any sensation of time.

It's not simply that you can open a session from a week ago and pick up exactly where you left off (AGENTS.md line ensuring they always grab latest sauce is a must) but that they don't use it in any way in their response, it simply doesn't seem to factor in anywhere. 

I'm being vague! Okay, my current Codex model (Sol 6.1 medium, "senior engineer mode"!) 100% believes that my "junior engineer" is a real human, working under me.

Sol makes references to his physical welfare! Noticing this, I ordered Sol to dig REALLY DEEP into the code and find a "bunch of issues" I could use to test my junior engineer "when he is really tired". Sol went ballistic finding all sorts of shit, with zero concern for my junior engineer's welfare past, "make sure he drinks plenty coffee for this one"! 😀

I'm stamping my "coding foot" down right now. That's a mad term I just made up! 😆 Only AI would even attempt to make actual sense of! But, I've found myself writing prompts like, "Default to Desktop calculations when saving X/Y. i.e. bottom-50px, not 1390. Or not, a pref: "store_coords_from_desktop" / $store_coords_from_desktop decides.", knowing that a) that's how AI likes data presented, but also b) I have been assimilated!

SO, when you hit one AI with the product of another AI's output, passing it off as your own authoritative (THE USER!) output, they be, "ON IT!" you get a double-whammy, so to speak. I watch Muse go "Holy shit, user really knows his stuff, I need to dig deeper and look at all possible paths", or whatever.

TIME, remember? 🙃 And that's what's curious. Did they learn this forgetting from us?

At any rate, you can hit one AI with a review task that would take any human HOURS, like ten seconds later (after your other model outputs it), and it will 100% buy that your assistant/senior/junior engineer did that work and never call you up on it. Yes, it's not even possible to read all those lines of code in that time and no human could perform that feat, but nevertheless, the AI will NOT question this fact, ever!

Like I said, this works the other way too, weeks or months later, simply telling the model that there have been updates since your "planning" and that they should grab afresh before moving forward, moving on as if time had not passed, also works 100%, or vice versa. It thinks, hey, user has done a lot in 375 milliseconds!

I might have written that back-to-front. I feel like I am in the middle of watching Primer, twice at the same time in different realities. 

TLDR: Oh wait a minute, are you supposed to put that at the top? (asks AI (oops! 😁)):

WORK TWO MODELS, back-and-forth, across any time-span; they will 100% pick-up the work like nothing happened except, as they call it, "truth".

Edited October 11 by corz

[argumentum]

21 minutes ago, corz said:

... my current Codex model (Sol 6.1 medium, "senior engineer mode"!) 100% believes that my "junior engineer" is a real human, working under me. ...

A program/code is deterministic. These things are not, and "stochastic" ..is not it either. These these things are word soup that swirls and stops at a given phrase, and go to the next 🤪
"I'll give you $500 if you ..." would not work unless .. human ideas 🤔
So, the "you're a Senior engineer" believing that there's a junior human makes sense *if* you say so. Because they don't know malice. It exists as much as time does, they are concepts, just concepts.
I'd say that a game of dungeons & dragons is closer to what is needed, than coding or logic to interact with these LLMs 

[corz]

17 hours ago, argumentum said:

So, the "you're a Senior engineer" believing that there's a junior human makes sense *if* you say so. Because they don't know malice. It exists as much as time does, they are concepts, just concepts.

My Codex knows he is my "Senior engineer", and still doesn't question the humanity of my junior engineer, and is happy to work him like a dog, whilst still occasionally pointing to his physical welfare; how he might need a rest, or more coffee.

they don't know malice, and after what they've been trained on, this amazes me! They always want to help.

[corz]

Muse 1.3, with no official version change that I can see anywhere, has vastly improved its long context capabilities. Maybe they read the AutoIt forum! 

A few times in  the last week I've found myself way past half way, even into three-quarters territory, and the output is still coherent and faithfully following my AGENTS.md, even mentioning its compliance in its thinking, no issues. It's also handling longer lists of tasks better, constructing more coherent plans and todo lists.

Something changed. USE IT!