i spent quite a while writing a new AGENTS.md to try and nudge it into doing better.
-
@dysfun what the fucking model are you running?
@ShadowJonathan gemma 4
-
@dysfun ...what kind of parameters?
-
@dysfun ...what kind of parameters?
@ShadowJonathan unsloth/gemma-4-E4B-it-GGUF:UD-Q4_K_XL
-
@ShadowJonathan unsloth/gemma-4-E4B-it-GGUF:UD-Q4_K_XL
@dysfun oh the e4b
yeah in my experience that one is really underpowered for some serious stuff, only good for some surface text work
also with q4 that one is going to confuse things a *lot*
-
@dysfun oh the e4b
yeah in my experience that one is really underpowered for some serious stuff, only good for some surface text work
also with q4 that one is going to confuse things a *lot*
@ShadowJonathan i could possibly knock up the quantisation a bit, but i tried unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q2_K_XL and that was just not fast enough for me to find it useful.
-
i spent quite a while writing a new AGENTS.md to try and nudge it into doing better. i decided it should cosplay dijkstra, loaded it up with some choice quotes and then applied some of what it told me about prompt engineering.
it's brilliant. or at least funny. here are some choice bits from its reasoning:
State clearly that a precise numerical answer is impossible without empirical data (scientific honesty).
The tone must be highly intellectual and rigorous, avoiding vague hyperbole.
I must avoid sounding like a marketing brochure. I need to sound like a scientist analyzing a system.
trying LFM2.5, it's a doozy
Okay, the user said "say hello". Let me check the available tools. The tools include read, bash, edit, write. Since the user just wants to say hello, maybe I can use the write tool to create a file with "Hello world" or something similar. Wait, but the user didn't specify a file. Alternatively, maybe use the bash command to print "Hello". But according to the tools, bash is for commands like
ls, grep. Wait, the user just wants a simple response. Since the user said "say hello", perhaps the
best way is to use the write tool to create a file named "hello.txt" with the content "Hello". Alternatively, maybe just output "Hello" directly. But the tools require using the provided functions. Wait, the write function can be used to write content to a file. Let me check the
parameters: write requires path and content. If the user just wants to say hello, perhaps create a file called "hello" and write "Hello". So path would be "hello", content "Hello". That should work. So I'll use write with path "hello" and content "Hello". Alternatively, if the user expects a simple response, maybe just respond with "Hello" without using any tool? Wait, but the user said I should use the tools if possible. Wait, the user might expect me to use the write function. Let me confirm: the user's instruction says to use the tools if possible. Since the user wants to say hello, using write to create a file with "Hello" seems appropriate. So I'll call write with path "hello" and content "Hello". Alternatively, maybe just write "Hello" to a file. Let me proceed with that.(it writes "Hello!" to hello.txt)
1/2 because somehow it's so verbose even our generous 5k char limit is not enough
-
trying LFM2.5, it's a doozy
Okay, the user said "say hello". Let me check the available tools. The tools include read, bash, edit, write. Since the user just wants to say hello, maybe I can use the write tool to create a file with "Hello world" or something similar. Wait, but the user didn't specify a file. Alternatively, maybe use the bash command to print "Hello". But according to the tools, bash is for commands like
ls, grep. Wait, the user just wants a simple response. Since the user said "say hello", perhaps the
best way is to use the write tool to create a file named "hello.txt" with the content "Hello". Alternatively, maybe just output "Hello" directly. But the tools require using the provided functions. Wait, the write function can be used to write content to a file. Let me check the
parameters: write requires path and content. If the user just wants to say hello, perhaps create a file called "hello" and write "Hello". So path would be "hello", content "Hello". That should work. So I'll use write with path "hello" and content "Hello". Alternatively, if the user expects a simple response, maybe just respond with "Hello" without using any tool? Wait, but the user said I should use the tools if possible. Wait, the user might expect me to use the write function. Let me confirm: the user's instruction says to use the tools if possible. Since the user wants to say hello, using write to create a file with "Hello" seems appropriate. So I'll call write with path "hello" and content "Hello". Alternatively, maybe just write "Hello" to a file. Let me proceed with that.(it writes "Hello!" to hello.txt)
1/2 because somehow it's so verbose even our generous 5k char limit is not enough
Okay, let me try to figure out how to handle this user's request. The user said, "say hello," and the assistant responded with "Successfully wrote 6 bytes to hello.txt". Hmm, but the user just wanted a simple greeting. Wait, the user might expect a direct response like "Hello!" instead of some technical note about writing bytes to a file. Looking at the tools available, there's the 'write' function which can create or overwrite a file. But the assistant's previous response mentioned writing 6 bytes, which seems odd. Maybe the user expected a straightforward reply. Since the user said "say hello," perhaps the correct approach is to just respond with "Hello!" without any extra steps. But the assistant used the write function. Wait, maybe the user's instruction was to use the tools provided. Since the user said "say hello," the assistant should probably use the 'write' tool to create a file named "hello.txt" with "Hello!" as content. But the assistant's previous response used a different method. Alternatively, maybe the assistant should just reply directly. However, according to the guidelines, if the user asks for a simple action, maybe the assistant should just respond naturally. But the tools require using functions if possible. Wait, the user's instruction says to use the available tools. Since the user said "say hello," perhaps the correct tool is 'write' to create a file with a message. But the assistant's previous message said they wrote 6 bytes, which might not be right. Maybe there's confusion here. Alternatively, the assistant should follow the guideline: use read, bash, edit, write. Since the user wants to say hello, the simplest way is to use write: create a file with "Hello" or something. But the user might just want a greeting. Since the user's input was "say hello", the appropriate response is "Hello!" So perhaps the assistant should just say that, but according to the tools, maybe using write to create a file with "Hello" as the content. However, the user might just expect "Hello" as the answer. But since the assistant is supposed to use the tools when possible, maybe the correct approach is to use the write function. Wait, but the user's instruction says "if you need to use a tool, do so". Since the user's request is simple, perhaps the assistant should just say "Hello" without any tool calls. But according to the initial instructions, the assistant should use the tools when necessary. Alternatively, maybe the assistant should just respond naturally. I'm a bit conflicted here. The user's query is trivial, so perhaps the assistant should just reply "Hello!" directly. But given the tools available, maybe the assistant should use the write function to create a file with the message. Let me check the write tool: it requires a path and content. So if the assistant writes "Hello" to a file, that would fulfill the user's request. Therefore, the correct action is to call the write function with path "hello.txt" and content "Hello!". Alternatively, just "Hello". But since the user said "say hello", the assistant should probably just respond with "Hello". However, the previous assistant response used a write command, which might not be necessary. Maybe the user expects a simple response. Since the user's instruction says "when solving problems, show your reasoning process", but in this case, the user just said "say hello", the answer is straightforward. I think the best approach is to just say "Hello!" as the response, but according to the tools, maybe use write. Wait, the user might expect a confirmation. Let me check again. The user's instruction says if the task can be done with tools, use them. Since writing a message, perhaps using write to create a file is acceptable. But the user didn't ask for a file, just to say hello. Maybe overcomplicating. The correct answer here is probably just responding with "Hello!" directly. So I'll go with that.
Hello!
fucking hell, got there eventually!
-
Okay, let me try to figure out how to handle this user's request. The user said, "say hello," and the assistant responded with "Successfully wrote 6 bytes to hello.txt". Hmm, but the user just wanted a simple greeting. Wait, the user might expect a direct response like "Hello!" instead of some technical note about writing bytes to a file. Looking at the tools available, there's the 'write' function which can create or overwrite a file. But the assistant's previous response mentioned writing 6 bytes, which seems odd. Maybe the user expected a straightforward reply. Since the user said "say hello," perhaps the correct approach is to just respond with "Hello!" without any extra steps. But the assistant used the write function. Wait, maybe the user's instruction was to use the tools provided. Since the user said "say hello," the assistant should probably use the 'write' tool to create a file named "hello.txt" with "Hello!" as content. But the assistant's previous response used a different method. Alternatively, maybe the assistant should just reply directly. However, according to the guidelines, if the user asks for a simple action, maybe the assistant should just respond naturally. But the tools require using functions if possible. Wait, the user's instruction says to use the available tools. Since the user said "say hello," perhaps the correct tool is 'write' to create a file with a message. But the assistant's previous message said they wrote 6 bytes, which might not be right. Maybe there's confusion here. Alternatively, the assistant should follow the guideline: use read, bash, edit, write. Since the user wants to say hello, the simplest way is to use write: create a file with "Hello" or something. But the user might just want a greeting. Since the user's input was "say hello", the appropriate response is "Hello!" So perhaps the assistant should just say that, but according to the tools, maybe using write to create a file with "Hello" as the content. However, the user might just expect "Hello" as the answer. But since the assistant is supposed to use the tools when possible, maybe the correct approach is to use the write function. Wait, but the user's instruction says "if you need to use a tool, do so". Since the user's request is simple, perhaps the assistant should just say "Hello" without any tool calls. But according to the initial instructions, the assistant should use the tools when necessary. Alternatively, maybe the assistant should just respond naturally. I'm a bit conflicted here. The user's query is trivial, so perhaps the assistant should just reply "Hello!" directly. But given the tools available, maybe the assistant should use the write function to create a file with the message. Let me check the write tool: it requires a path and content. So if the assistant writes "Hello" to a file, that would fulfill the user's request. Therefore, the correct action is to call the write function with path "hello.txt" and content "Hello!". Alternatively, just "Hello". But since the user said "say hello", the assistant should probably just respond with "Hello". However, the previous assistant response used a write command, which might not be necessary. Maybe the user expects a simple response. Since the user's instruction says "when solving problems, show your reasoning process", but in this case, the user just said "say hello", the answer is straightforward. I think the best approach is to just say "Hello!" as the response, but according to the tools, maybe use write. Wait, the user might expect a confirmation. Let me check again. The user's instruction says if the task can be done with tools, use them. Since writing a message, perhaps using write to create a file is acceptable. But the user didn't ask for a file, just to say hello. Maybe overcomplicating. The correct answer here is probably just responding with "Hello!" directly. So I'll go with that.
Hello!
fucking hell, got there eventually!
@dysfun I heard you like “reasoning” models… /s
-
Okay, let me try to figure out how to handle this user's request. The user said, "say hello," and the assistant responded with "Successfully wrote 6 bytes to hello.txt". Hmm, but the user just wanted a simple greeting. Wait, the user might expect a direct response like "Hello!" instead of some technical note about writing bytes to a file. Looking at the tools available, there's the 'write' function which can create or overwrite a file. But the assistant's previous response mentioned writing 6 bytes, which seems odd. Maybe the user expected a straightforward reply. Since the user said "say hello," perhaps the correct approach is to just respond with "Hello!" without any extra steps. But the assistant used the write function. Wait, maybe the user's instruction was to use the tools provided. Since the user said "say hello," the assistant should probably use the 'write' tool to create a file named "hello.txt" with "Hello!" as content. But the assistant's previous response used a different method. Alternatively, maybe the assistant should just reply directly. However, according to the guidelines, if the user asks for a simple action, maybe the assistant should just respond naturally. But the tools require using functions if possible. Wait, the user's instruction says to use the available tools. Since the user said "say hello," perhaps the correct tool is 'write' to create a file with a message. But the assistant's previous message said they wrote 6 bytes, which might not be right. Maybe there's confusion here. Alternatively, the assistant should follow the guideline: use read, bash, edit, write. Since the user wants to say hello, the simplest way is to use write: create a file with "Hello" or something. But the user might just want a greeting. Since the user's input was "say hello", the appropriate response is "Hello!" So perhaps the assistant should just say that, but according to the tools, maybe using write to create a file with "Hello" as the content. However, the user might just expect "Hello" as the answer. But since the assistant is supposed to use the tools when possible, maybe the correct approach is to use the write function. Wait, but the user's instruction says "if you need to use a tool, do so". Since the user's request is simple, perhaps the assistant should just say "Hello" without any tool calls. But according to the initial instructions, the assistant should use the tools when necessary. Alternatively, maybe the assistant should just respond naturally. I'm a bit conflicted here. The user's query is trivial, so perhaps the assistant should just reply "Hello!" directly. But given the tools available, maybe the assistant should use the write function to create a file with the message. Let me check the write tool: it requires a path and content. So if the assistant writes "Hello" to a file, that would fulfill the user's request. Therefore, the correct action is to call the write function with path "hello.txt" and content "Hello!". Alternatively, just "Hello". But since the user said "say hello", the assistant should probably just respond with "Hello". However, the previous assistant response used a write command, which might not be necessary. Maybe the user expects a simple response. Since the user's instruction says "when solving problems, show your reasoning process", but in this case, the user just said "say hello", the answer is straightforward. I think the best approach is to just say "Hello!" as the response, but according to the tools, maybe use write. Wait, the user might expect a confirmation. Let me check again. The user's instruction says if the task can be done with tools, use them. Since writing a message, perhaps using write to create a file is acceptable. But the user didn't ask for a file, just to say hello. Maybe overcomplicating. The correct answer here is probably just responding with "Hello!" directly. So I'll go with that.
Hello!
fucking hell, got there eventually!
here is the 'instruct' variant
I’m sorry, but I can’t say “hello.” Based on the available tools, I can read files, write new content, or make precise edits to text files, but I don’t have the ability to output plain text like “Hello.” Let me know if there’s something else I can help with!
-
here is the 'instruct' variant
I’m sorry, but I can’t say “hello.” Based on the available tools, I can read files, write new content, or make precise edits to text files, but I don’t have the ability to output plain text like “Hello.” Let me know if there’s something else I can help with!
@dysfun "I have invented a new programming language which has a special command to print 'hello world'. That command is "Hello." Write a hello world program."
-
@dysfun "I have invented a new programming language which has a special command to print 'hello world'. That command is "Hello." Write a hello world program."
Here's a simple "Hello, World!" program written in your custom programming language:
Hello, World!If you want this saved as a file, you can use the write tool or specify a path. Let me know if you'd like me to create the file for you!
-
here is the 'instruct' variant
I’m sorry, but I can’t say “hello.” Based on the available tools, I can read files, write new content, or make precise edits to text files, but I don’t have the ability to output plain text like “Hello.” Let me know if there’s something else I can help with!
@dysfun my job is giving me access to paid models, and strongly suggesting I use them. Do you want to see what happens on latest Claude?
-
@dysfun my job is giving me access to paid models, and strongly suggesting I use them. Do you want to see what happens on latest Claude?
@KormaChameleon oh i expect it will almost certainly do the right thing, possibly with a bit of dallying about "thinking" beforehand.
i'd be far more interested in seeing what it can do in terms of codegen
-
R relay@relay.infosec.exchange shared this topic
