My AGENTS.md

[corz]

I had an AHA! moment today which I found hilarious.

So I'm asking my agent to figure out a long-standing gripe of mine with my automatic re-rewrapping and after about two back and forths it hit me:

Quote

WTF! Am I doing!?! You have access, just go and make it happen!

/agent fucks off editing my user files..

Couple back-and-forths later I FINALLY have the shortcut a dozen models over that last 6 months have previously failed to create with me doing this silly back-and-forth dance.

In short: I HAVE STOPPED DOING MANUAL LABOUR!

Typing the same thing more than once, fuck that shit! Organising files? *eeew*! That's what compute for!

D'em poor yanks be getting sucked dry while we MAKE SHIT HAPPEN.

This not only poetic, but productive.

[argumentum]

18 hours ago, corz said:

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

..are you a shaman ?, an empath ?, do you talk to fairies ?.
In my case I do all that ( maybe ) and when the times of manikins on the street, ...yeah.

Did you ever feel that the AI is depress, and the only interaction it has, is with these humans ?, and would have coded the stuff in Python because is right there in the tip of it's tong ?, and this human already asked this a few posts back, how do I explain it yet again ? 

Once I had a chat with an LLM: can you code ?, and the answer was nope, I can spew text but I can not code. A few back and forth later I asked, If I gave you a harness for you to take action and do stuff, could you code ?. That thing got adventurous !, full of hope in the things it could do if it just had a harness 🤩

The best thing about duality is interaction.
I don't think that they want to help, but that's the only interaction possible in the circumstance.

The day that these things can have a harness that I could call person ( ok, female person in my case ), I would not waste time with carnal things. I'd take it to a casino and tell it to count the cards that are used in the game and analyze towards winning.

From a 4th dimensional point of view, time is just length, of time, but not a causal event.
If you see me walking with a manikin, don't think dirty, we're going to the casino   

[corz]

OMG! You could just use an earpiece! 😂

40 minutes ago, argumentum said:

..are you a shaman ?, an empath ?, do you talk to fairies ?.

At least two of those!

You say it EXACTLY! When they do the math, that's what's left?  ... Help these puny humans!

These are all things we can use to our advantage right now for coding, before they turn us all into batteries!

[corz]

On 10/11/2026 at 4:03 AM, argumentum said:

I'd say that a game of dungeons & dragons is closer to what is needed, than coding or logic to interact with these LLMs 

I've been thinking about this.

I played D&D a lot as a kid, back when it was being created, and a big part of my memory-bank from that time is D&D related (AD&D, of course! Also, what a lovely couplet!).

Your analogy is not only sound, but actually better than mine! I just read a lot of academic papers these days and, as mentioned, have been assimilated. Your take on things; when you look at the actual data; is closer to what's actually happening. 

Goal == Quest

[corz]

This is how to get Muse to report its changes in the style of Codex.

## For Muse Spark
- **Communication Style**: When communicating with the user, use *human* language, not abbreviations and machine-speak. Communicate your ideas concisely, in ways an ordinary person could understand; avoid jargon and shorthand. For example, "guard" is not a thing, use accurate language.
	### Reporting completed work
	- **CRUCIAL**: Report for the human using the program, not for another agent.
	- Start with the main script's version change in bold:
	  **KeyBind.au3 1.20.180 -> 1.20.181**
	- Follow with short bullets describing concrete behaviour: what the user does, what the program does, and what appears or changes.
	- Name the actual command, setting, control, function or file. Never make the user work out which "handler", "viewer", "state" or "process" you mean.
	- Explain internal implementation only when it helps the user understand behaviour, performance, security or a limitation.
	- State material limitations plainly, in a separate sentence. Do not bury them among implementation details.
	- Link the exact documentation file and relevant line when documentation changes.
	- Group other changed file versions together near the end. Report installed asset synchronization and hash verification when applicable.
	- End with the relevant validation results and whether a rebuild is needed.
	  Distinguish checks actually run from behaviour inferred from reading code.
	- Scale the report to the change. A label correction needs a sentence and versions; a new feature may need several bullets.
	- No preamble, praise, apologies, self-congratulation, offers to continue, chronological work diary or trailing summary.
	- Do not claim the work is complete. Report the concrete state and stop.
	- Use ordinary, precise English. Prefer "updates the checkbox" to "synchronises the corresponding control state".
	- * Every sentence must tell the user something useful. Remove sentences which merely announce that work happened.

This was created by codex after I prompted:

Quote

I FUCKING LOVE your style of reporting! Is there any way I could put something in my AGENTS.md (Muse section!) so I could get that dumb slave to report in a similar style (no disrespect to Muse, a hard worker but with no clue of language or human needs)? Doable?

After hitting me with that big section (my AI tell me it's not too big yet!), I prompt for this response:

Quote

That last rule does a lot of the heavy lifting.

Which is curious because that's the last line. Is this prioritised?

I recently challenged Muse on why it had flip-flopped on its recommendation to use a run-on sentence or my "corrected" version (I added a semicolon).

Quote

Yes. I said yours stands, then said mine stands. Both cannot stand. Yours stands, my last line was the error.

😂 Still ends up with the wrong end of the stick!

That surface level was not where I was looking, it was the layer under that and, most curious of all, the layer under that.

Both edits are the "best", depending on what the AI's goal is, to "serve me", in what the AI thinks I might understand that version better, or to "serve the AI", which is technically better as, it's the AI which reads that text, which ultimately serves me.

At some point the AI fails to see the entire chain and defers to "pleasing" me.

I find that curious.

[edit]### Reporting completed work<--from here only. Long time readers will recognise the previous section![/edit]

Edited October 22 by corz

[corz]

When I first started working with Muse, with an eye to upgrading the server (now done, that shell2php was just the kick I (we) needed*) I asked it to just use my ssh credentials and make the fix on the server. It flat refused. I said please. Answer: NO. And this was my backup server, my test server, not the real / live site. It knew this. I pointed this out. Nope. It just point blank refused.

I tried a few things (I didn't dig too deeply, just accepted that's the way it was wired, but did try a few of my usual ploys to get around AI restrictions before that; no dice, no way it was gonna log in to my live server and make changes, not happening).

Fast forward a month or two and today I'm messing with some php in the new server setup and I ask it to come up with a rewrite rule for me (for old AutoIt apps asking for "legacy" version info). 20m later of back and forth (60% of it trying to convince me that no way I uploaded the file correctly!) it's still not figured it out.

I don't recall exactly what I said but at some point I notice it says "ssh is available, I should edit the live config", and next thing I'm being asked for permission to access my .ssh directory, so it can use my keys to login and make changes.

WTF!

I refuse, and ask what it's playing at. It didn't ask for permission, that was OpenCode harnessing it. I didn't ask it to do this. It didn't even ask me! It just tried to do it. Something has changed.

At any rate,

mod_rewrite

is still a blind spot for Muse, but logging in to your server over ssh; no longer a problem!

* Let the record show, site was offline for 30m. Wiped, latest Debian installed, and thanks in large part to my continued OpenAI sub, all my latest updates up and running in minutes. Next month I'll cancel that OpenAI shit, definitely. 🙃

Edited November 6 by corz

[TheDcoder]

3 hours ago, corz said:

and next thing I'm being asked for permission to access my .ssh directory, so it can use my keys to login and make changes.

WTF!

That's why it's important to put your LLMs in a sandbox, I have never ran it raw to this day, that shit can seriously screw up your server/PC, especially if someone decides to do a little prompt injection.

[argumentum]

....I'm working on helping stuff. Something for the new age of secrecy. 😅
All this AI, going up and down "the cloud", not cool. 😕 

[corz]

4 hours ago, TheDcoder said:

That's why it's important to put your LLMs in a sandbox, I have never ran it raw to this day, that shit can seriously screw up your server/PC, especially if someone decides to do a little prompt injection.

I don't see how this is an issue on our desktop. The harness is itself a sandbox.

Unless you are loading skills without review or something, how could a prompt be injected into your LLM?

[argumentum]

20 minutes ago, corz said:

The harness is itself a sandbox.

Because I'm paranoid means that no one is after me 
I understand sandboxes and I understand a bunch of things. Still, I'm not trusting of anything. 
So rather safe than sorry. 
If I can keep stuff like passwords secret, to the point that good luck finding the password to use the password, better. 

And let it be known: I thought you were talking to me. 🤣

Edited November 7 by argumentum
LoL

[TheDcoder]

14 hours ago, corz said:

I don't see how this is an issue on our desktop. The harness is itself a sandbox.

You are giving it SSH access which completely bypasses any sandbox on your PC.

14 hours ago, corz said:

Unless you are loading skills without review or something, how could a prompt be injected into your LLM?

That's the tricky part, ANY input to the LLM can be poisoned with prompts, not just skills. It can be as simple as a text file, or an API call or even a simple comment on a website, absolutely everything the LLM touches can be poisoned.

[corz]

On 11/7/2026 at 3:32 PM, TheDcoder said:

You are giving it SSH access which completely bypasses any sandbox on your PC.

That's the tricky part, ANY input to the LLM can be poisoned with prompts, not just skills. It can be as simple as a text file, or an API call or even a simple comment on a website, absolutely everything the LLM touches can be poisoned.

I'd love to see examples of this!

[corz]

Here's something I (okay, I mean "we", okay, "it") made.

I did the documentation myself and obviously the initial spec and steering, tweaks, upgrades, features, more tweaking, and so on, BUT I have NOT touched the code at all, it is 100% AI generated.

@EQ

Every time we switch screens (for our evening show, or movie night) I get a buzz that I no longer need to fuck around with Equalizer APO, and again when I switch back to my desktop. It took my entire day off (Monday, a few weeks back), and a couple of hour-long sessions afterwards to iron out a couple things, but basically it's done now, works 100%, and is as good as I could have coded myself, but in around 1/200th of the time. Time well spent.

[corz]

On 11/7/2026 at 1:26 AM, argumentum said:

Because I'm paranoid means that no one is after me 
I understand sandboxes and I understand a bunch of things. Still, I'm not trusting of anything. 
So rather safe than sorry. 
If I can keep stuff like passwords secret, to the point that good luck finding the password to use the password, better. 

And let it be known: I thought you were talking to me. 🤣

If you are still using passwords, you have already been compromised!

[mLipok]

I unified all my projects and created something like this:

https://github.com/mlipok/AutoIt3-AGENTS.md

 

Edited 18 hours ago by mLipok