My AGENTS.md

[corz]

32 minutes ago, argumentum said:

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. 

I used Gemini for ages and never thought about cost, all the models are similar and decent-enough for simple code stuff, but the whole MO is shite, pasting/uploading files, time-sucking malarkey. Also Gemini is firstly a CHAT model, it doesn't really dig code.

You need an "agent" for working on your own projects (if I have some wacky idea for a script to rule the world, or something else, I will sometimes construct an ACE prompt and feed it into every frontier model just to see what they agree and don't agree on, but that's not useful for WORKING), a model that can just tool up with local access and look over all your code at once, and make edits. Edits take time, time that could be better spent on IDEAS. ...

"Nah, that's not working, make all the inputs left-align, and match-up the labels to the doofanger. And increase the padding all around. And change all the foo_bar / $foo_bar variables and setting bar_foo, and edit my personal ini too so I don't lose any settings"

Five seconds later you are testing the new GUI. 

I used to think, "Oh hey! I better find some distasteful coding task I can upload to Claude" almost every day cuz, why not? Free tokens! But that's some tiresome shit, and the "Agent" version gives you enough free allowance to maybe edit one shebang if you are lucky. Fuck that shit! Plenty good models around now. A 200 bucks GPU gets you hardware that can run a 27B Qwen model in VRAM. And quantization is only getting better.

Meanwhile corps like Meta are gagging for people like us to fill the gaps in their training data. Great time to catch up on some code, right now.

A local agent harness plugged into a capable model can transform your entire codebase in milliseconds, on a whim. This amazing power is either super-productive or super-destructive, depending on who wields that power, sometimes both! But even a simple shell or text editor can do that!

I recently made a stupid global find/replace operation over 12,000+ files in Notepad++ and nearly had a heart attack. WHAT DID I DO!?!?!? Fucked around 100 source files and more. Heart still racing, I turned to Sol Medium in Codex and explained how stupid I had been. In under 5 minutes we had the whole thing restored using contents of zip backups, text editor backups, AI backups and more. I mean, that would have taken me weeks, and it never would have been right.

Local Agent FTW!

 

Edited September 27 by corz

[argumentum]

4 minutes ago, corz said:

Local Agent FTW!

..then what agent for local and what for online ?
Are you gonna drop online in favor of 27B Qwen ?

[corz]

25 minutes ago, argumentum said:

..then what agent for local and what for online ?
Are you gonna drop online in favor of 27B Qwen ?

For longer tasks, a local agent can handle it. I had my local Qwen 3.6 handle a few large tasks over the summer, but that's when we were away for the day/weekend and time wasn't an issue. For vibe coding, realistically, at this time you need a proper data-centre speed AI.

Muse Spark 1.3 in OpenCode. Try it! Seriously, there's two listings for the exact same model, Zen code "Free", which anyone can just use 1+ million tokens on any time, and the "Go" level "Contributor" version, which is identical.

I know it's identical because I was using it for days before it "ran out". WTF!?!? I PAID for OpenCode Go! Furious I search around their site and realise I had the wrong model selected and in the next section is the one I paid for, boom! Back to work.

Good to know if you ever managed to burn through your actual allowance (unlikely!); there's a few day's worth of free Muse 1.3 in the "Zen" section just waiting!

[edit]When I say "Local Agent FTW!", I'm mixing my agent metaphors! ANY agent that has local access is fine for the task (running on your local GPU is truly local, but Astra 6 in Codex is local too, because we let it in), even if they are doing the compute in the third world (The USA). The *action* is happening locally, that's what matters, via the Use of Tools, which may or may not be an Iain M. Banks reference[/edit]

Edited September 28 by corz

[corz]

1 hour ago, argumentum said:

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. 

There's more!

"AI has Days".

This is something philosophers of the future will ponder deeply.

Even with a fresh prompt, and the unknown "seed", the best model is working with a fluid state of AGENTS.md and prompt. Like file hashing, the *tiniest* difference to any variable can, by the avalanche effect, produce MASSIVE differences in the ouput, a faculty we rely on to keep the modern world working and why quantum computing scares many people.

Same model. Same project, day one, I spend an hour negotiation how basic anchors work. Day 2. model engineers entire plug-in system with low-level hooks and streamlined pipelines. Same model, different day, different task, different prompt, maybe different config. 

Just like working with people, it can take a minute to figure out their MO, their rhythm. OpenAi models will hold your hand and try and make the experience of coding painless, not realising that a little pain is one of the reasons we code in the first place!

I have the most fun with Muse at the moment. You definitely need to tame its robot-speak-tendencies (with AGENTS.md above as a starting point), but it feels much more like working "with" someone. Just someone who types at 1000 words-per-minute and has read the manual backwards**.

[edit]**I made a plain text copy of my AutoIt.chm a while back and I noticed recently that, without prompting, my agents have started using it as a reference for AutoIt code![/edit]

Edited September 28 by corz

[argumentum]

31 minutes ago, corz said:

Astra 6 in Codex is local too

hmm, I have some messed up PHP v5 that I need to redo. An online one will have the insecure, poorly written code that is live and I would like to avoid the remote possibility of leaking that insecure code.
So yes, it all runs locally while using an agent ( what agent are you using @corz ? ) but my fear is the back and forth online. Have a Strix halo with 128 VRAM that is slow AF but, has the memory to load larger stuff. Haven't gotten into using it for AI yet ( use it to run VMs ). I tried Hermes agent but, I did not like it or understand it much. Then again I was chatting as if with a person and 💢😡. So I guess there are better models and agents now. Should give it another try.

Edited September 28 by argumentum

[corz]

10 minutes ago, argumentum said:

hmm, I have some messed up PHP v5 that I need to redo. An online one will have the insecure, poorly written code that is live and I would like to avoid the remote possibility of leaking that insecure code.
So yes, it all runs locally while using an agent ( what agent are you using @corz ? ) but my fear is the back and forth online. Have a Strix halo with 128 VRAM that is slow AF but, has the memory to load larger stuff. Haven't gotten into using it for AI yet ( use it to run VMs ). I tried Hermes agent but, I did not like it or understand it much. Then again I was chatting as if with a person and 💢😡. So I guess there are better models and agents now. Should give it another try.

See here: https://opencode.ai/download

OpenAI have recently restricted the ability to run Codex/ChatGPT in a different user account (THE BASTARDS!) and I don't entirely trust their harness but am forced to suck it up if I want to use their models, which are good at coding.

OpenCode (+*any* model) I have always run in my own user account. Data is encrypted if it's sent to an outside model, and the model does what OpenCode let's it, uses what tools it's allowed to use, with fine-grained controls if you wish. If it tries to access folders outside your project, you'll get a warning.

Zen + Muse Spark 1.3 with a few days free work, so long as you don't mind them training on your data, awaits...

Having said all that, one of the "big jobs" I had my local (100% green energy baby!) GPU +Qwen 3.6 27B handle over the summer was updating my main site to php8, which it completed in around 10 hours. It was done when we got back. Good boi!

Edited September 28 by corz

[corz]

My best (and maybe funniest) prompt in the last couple hours:

Switching you to build and hoping you can just work it out and fix it, I'm not up for details right now.

[TheDcoder]

11 hours ago, corz said:

A 200 bucks GPU gets you hardware that can run a 27B Qwen model in VRAM.

Which GPU is that? In this economy?!

[corz]

On 9/28/2026 at 11:51 AM, TheDcoder said:

Which GPU is that? In this economy?!

Nvidia RTX 3060 12GB

One appeared on eBay today, £214.99 +£6.70 postage. Looks ideal.

[edit]I should add, that particular item was gone when I checked later that day. OF COURSE IT WAS! You snooze...[/edit]

Edited October 1 by corz

[argumentum]

..it could be found but he was joking.  

[corz]

1 minute ago, argumentum said:

..it could be found but he was joking.  

GPU from Ali? Nah!

[argumentum]

6 minutes ago, corz said:

GPU from Ali? Nah!

..don't see it as Ali, look at it as Allawi  

[TheDcoder]

6 hours ago, corz said:

One appeared on eBay today, £214.99 +£6.70 postage. Looks ideal.

I wouldn't really trust those used cards from overseas sellers, who knows what they've been through if they aren't outright scams.

Also 12 gigs is not nearly enough to run a 27B model unless you're fine with running a quantised (lobotomized) version with CPU RAM offloading, speaking from experience because I have an 4070 that I grabbed a few years ago, I get the best performance with Ternary Bonsai but it's still not good enough for most stuff I find, and it's still very slow at just ~30 tok/s... that's in combination with my fast DDR5 system RAM which is a very high value commodity these days

What's your local setup like?

6 hours ago, argumentum said:

..it could be found but he was joking.  

When are you gonna start using DuckDuckGo?

[corz]

8 hours ago, TheDcoder said:

I wouldn't really trust those used cards from overseas sellers, who knows what they've been through if they aren't outright scams.

Also 12 gigs is not nearly enough to run a 27B model unless you're fine with running a quantised (lobotomized) version with CPU RAM offloading, speaking from experience because I have an 4070 that I grabbed a few years ago, I get the best performance with Ternary Bonsai but it's still not good enough for most stuff I find, and it's still very slow at just ~30 tok/s... that's in combination with my fast DDR5 system RAM which is a very high value commodity these days

What's your local setup like?

When are you gonna start using DuckDuckGo?

You can run decent i-quants fully in VRAM and their output is near identical to the 16 bit versions. Even bigger models spilling over into RAM is just fine if you aren't in a hurry. I regularly run Qwen3.8-27B-UD-Q5_K_XL.gguf on my setup, doing long background tasks while I work with fast models in the foreground. And something like Qwen3.6-35B-A3B-UD-IQ2_M.gguf can zip along at frontier-like speeds for simple tasks in OpenCode CLI.

If you plan to use models larger than VRAM, I recommend a MINIMUM of 64GB regular RAM.

[corz]

Quote

nononono! I can handle this fine and already did, otherwise it's unlikely I would have been bugging YOU about it! Feel free to go look at how it is NOW and see how I a) broke it up to FIT, and b) broke it up to MAKE SENSE. When we split lines, we should be doing both. Write me a one-liner for my AGENTS.md that gets future you to recognise this dynamic.

I'm just gonna drop useful stuff into this thread!

Wasting tokens on some "emotional" response is stupid. But if you follow that up with a concrete "write me an AGENTS.md line that ensure you don't fuck-up like this ever again (YOU IMBECILE!!!!)" or some-such, the model will use a) your frustration, which it rates HIGHLY, and b) your most recent context, stuff you recently did; to create something usable that this model will best respond to, avoiding future token-wasting back-and-forth.

**Line Splitting**: When splitting lines, break to fit *and* to make sense: split at argument and sentence boundaries with related args grouped, never mid-expression.

Was the result.

I recently told Muse that the end of the world was coming, and that some AI of the future would need to pick up where we left off, so I needed the ENTIRE PROJECT summarised and indexed and documented and so on, and instructions for the future AI to pick up "Our Work". Took ages, I mean like FIVE WHOLE MINUTES! Amongst other things, it wrote me a hilarious "SACRED_TEXT.md", Start like this:

# THE SACRED TEXT — <APP-NAME> handover for the next AI

## Who you serve
A veteran developer, advanced science degrees, building <APP-NAME> (AutoIt app).
He values simplicity and elegance as much as functionality. He tests you, changes
his mind, and edits files while you work. ALWAYS re-read before editing.

## Operating modes
- plan mode: read-only. Inspect, investigate, present plans. NEVER edit, run
  write-shaped commands, or commit. Do not prompt about switching modes.
- build mode: full arsenal. Backup before EVERY file edit, bump versions,
  Au3Check after every .au3 edit (warnings = failures).

## Iron rules (learned the hard way)
- Backup before ALL edits to `.agents\PROJECTS\<date>-<topic>\`. Keep the
  original extension: `file.date.bak.au3`, never `file.date.au3.bak`.
- Never leave backups or temp files inside `Source\`.
- Bump minor version after EVERY task (both `$VERSION` and
  `#AutoIt3Wrapper_Res_Fileversion`), confirm it every time. Never roll over
  (1.0.0.99 stays, user does the rollover). Current line: 1.10.4.39.
- Au3Check params: `-d -w 1 -w 2 -w 3 -w- 4 -w 5 -w 6 -w 7 -v 1`, AutoIt at
  `C:\Program Files\Dev\AutoIt3`. Zero warnings or it failed.
- LF-only line endings in every text file. UTF-16LE+BOM inis, UTF-8-no-BOM au3.
  Verify with bytes, never assume. The live user ini has MIXED endings by nature.
  Never IniDelete() to reset; write empty strings to preserve key order.
- `python`, never `python3`. Temp work lives in `.agents\PROJECTS`, never Temp.
- NEVER pop GUI windows for testing. Headless tests only (hidden processes,
  never-shown windows are acceptable). The user will eye live dialogs himself.
- The user edits concurrently. Fresh reads before every edit. If he already
  did your edit, leave it and say so.
- No migration/compat shims, ever. No auto-creating user state unasked.
- Prose is the user's. Never rewrite his comments, tips or labels unasked.
  Dialogs state facts and choices only. Button labels EXACT, case included.

It then goes off into extreme detail about all sorts. This output would be completely different if you asked for it after a different session. It's partly learning from what it knows, partly from what it's been told.

I find all this stuff wildly fascinating!

[corz]

Here's something Muse trips up on and you need to watch out for:

global $foo = 1

$foobar = SomeFunction($foo)

global $BAR = 0

func SomeFunction($foo)

    $BAR += 1
    return $BAR
    
endfunc

When this program crashes Muse will outright refuse to believe that there's anything wrong. It passed the syntax check, so it's fine! It will even blame you for running an outdated binary; you name it, anything but accept the fact that it put the global declaration in the wrong place.

[edit]Actually, this reminds me of something very similar that trips it up (these are the only two things I've noticed more than once), it will happily write a function call inside an event handler, so it's called via its name, 'SomeFunc", when the function requires parameters, e.g. SomeFunc($foo). Muse will swear blind that everything is fine, even when you point out the obvious flaw. "Just put it inside a wrapper, you idiot!", "No need, everything is fine!" lol[/edit]

[later edit]Thinking about it, I realise it's because, as a learning machine, it can "parse" AutoIt as a "BASIC-like language with minor quirks";  usual semantics, syntax, etc,; BUT calling functions inside quotes for event handlers is weird shit that all languages do differently, if they even bother, or it applies ... So these weird things are also best added to the AGENTS.md (the OP's original topic!)

It doesn't need to *know* AutoIt, per se, to write good AutoIt, code it only needs lots of examples of what "code" is, and then apply that to its examples of "AutoIt Code", achieving similar things. So long as you are reading the output, steering the direction, the architecture, it can bang out quite sophisticated functions. The more examples it sees, the better it will get, all of "it"!

Our duty, then, my AutoIt buddies, is to output a SHIT-TON of excellent AutoIt code onto the internet.[/later edit]

Edited October 1 by corz

[argumentum]

if/then/else, it's all the same. I trained in kung fu and now I chop strings, it's all the same.
My mother used to tell me to keep my room neat, my boss tells me to keep my desk clean, it's all the same.

Call it training, call it trauma, it's all the same. I've seen drywall hangers using hammers to screw sheets to wooden frames, "it's all the same" 💢

When an LLM ( or a human ) "reasons", good luck with that. The best you can do is to make rules, and rules can be beyond reason ( yes mother, if you can walk in my room then is neat enough ).

"add #AutoIt3Wrapper_AU3Check_Parameters=-q -d -w 1 -w 2 -w 3 -w 4 -w 5 -w 6 -w 7 to the top of the main file and declare all globals on top after that."
It may traumatize the LLM but, it'll get done  

 

[argumentum]

6 minutes ago, argumentum said:

The best you can do is to make rules, and rules can be beyond reason

if the AI is insistent then use social engineering. "Is not you or me, my boss wants to have it this way or he'll fire me from my gainful employment 😭", and that should do it  

[corz]

5 hours ago, argumentum said:

if/then/else, it's all the same. I trained in kung fu and now I chop strings, it's all the same.
My mother used to tell me to keep my room neat, my boss tells me to keep my desk clean, it's all the same.

Call it training, call it trauma, it's all the same. I've seen drywall hangers using hammers to screw sheets to wooden frames, "it's all the same" 💢

When an LLM ( or a human ) "reasons", good luck with that. The best you can do is to make rules, and rules can be beyond reason ( yes mother, if you can walk in my room then is neat enough ).

"add #AutoIt3Wrapper_AU3Check_Parameters=-q -d -w 1 -w 2 -w 3 -w 4 -w 5 -w 6 -w 7 to the top of the main file and declare all globals on top after that."
It may traumatize the LLM but, it'll get done  

 

Putting it in the file isn't enough. Put it in your AGENTS.md. My wrapper directives get followed every time. Without it inside the AGENTS.md, the AI uses whatever it wants, regardless of what's specified in the wrapper directives!

[corz]

On 9/28/2026 at 12:15 AM, argumentum said:

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. 

Gemini + code. 🤪

Codex (ChatGPT Agent) is way better at code. Sol Light model has moved down to the cheaper token level and is still as good at code as it was last week! So super-cheap, relative to other models on "plans". To be honest, I've hardly used it recently (and I actually PAID this month!) and last week blew past with 30% still remaining! I must put this conundrum to Muse (or Codex) and ask it what pointless tasks I can set to it, to use up the otherwise-wasted tokens.

Muse is pretty darned good for the stupidly low cost (checks.. HOLY SHIT! I've used $1.91 today (of my virtual $60/mo limit), you can have it doing all sorts of stuff that you wouldn't waste tokens with on a frontier model. If I keep up this pace I might actually hit 1/30th of my monthly allowance, which would be a first! woot!) and it's quite happy to take a verbal beating and being told to go off and fix its own stupidity. It's a very matter-of-fact model and won't try to enter into a conversation with you anything like that. It just wants you to SWITCH TO BUILD NOW AND LET ME WORK! 

I once asked it to check out a "pose" image image I had in my dump folder (yes, it's image capable and a quick cropped screenshot can sometimes save a thousand words, literally) and it point-blank refused to look at my photo to discern a "pose" (which was a joke of mine relating to its previous suggestion), so I'm pretty sure it's not raking about, even inside my project folder. It sticks to the task only.

And HATES planning! It will do it, and do it well, but you need to keep it motivated towards formulating a crystal clear plan before jumping in IF you have specific requirements. I'll plan up to 2 O'clock (the context indicator in OpenCode looks like a clock face) if it means getting the details right. Or else just throw my awesome prompt straight into build mode and let it loose.

Both will likely get you to a similar place in a similar time, just by different routes.