A bunch of these should be enforce with linting, that way people who still hand-craft code get the same kind of feedback, e.g. Always use {}, even on a one-line "if" statement. & Keep function names short. Less than 30 characters.
Then this one really is a pattern that creates a lot of churn:
- Add a small, to the point, comment to explain what the block does and why. Use examples when possible. Propose ASCII drawings to explain complete systems.
I forbid my agents from adding any comments. I review the code and add comments manually. If I can't understand something despite having the context then I throw away the code instead of having an LLM generate comments to explain what it did. This way the code stays readable/debuggable by humans.
> This way the code stays readable/debuggable by humans.
Please take the following as expressed with genuine curiosity: Do you not use an editor with syntax highlighting and collapsible comments?
At least on JetBrains you can configure the editor to collapse all comments on open and to have the comments displayed in a low-contrast color. This way, LLMs add a bunch of comments, but it doesn't affect your actual experience in trying to read the code. If you encounter code that seems inexplicable, then and only then would you expand the comment to see if that helps you understand.
Comments should be written only when there is (hidden) complexity or external context strictly required. Otherwise it is just easier to read the code. Comments then signal one of two things: a) the following code is really complex and I need to tread carefully, or b) this code is complicated, and could benefit from a refactor.
In regards to agentic coding, all these comments are extra contents, driving down quality while increasing cost. Agents also tend to be inconsistent about updating comments, I've had cases repeatedly where a comment did not match the code, at which point it is just a documentation liability.
Also, the language model might not fully understand the code then add a comment, then the next iteration will treat assumptions in the comment as the truth.
Sometimes I try to add comments in a new session and the agent just don't have enough context for it to give a comprehensive sentence with full context on the why, then the agent will just describe what it does.
Human comment is in another level to answer the questions mainly like "why do it like this" for the later collaborators or the forget-ed self, so the important blocks live when it is needed and can be eliminated when it does not.
> If the human can read/understand it well enough to comment it, then it is readable by humans.
No, because the one who is writing the comment has context later reader dont. The writer knows what the requirements are, what he was trying to achieve and what he struggled to comprehend. Writer also presumably spent more time trying to understand it then the person coming later should.
It is not perfect, but the delta between 'cannot understand what agent wrote' and 'it makes sense to me right now' is already an improvement. That it may not be sufficient, doesn't mean it isn't a necessary condition.
Besides, it has always been like this. I sometimes can't even understand some of the things I wrote myself a couple of months ago, because I forgot the context. Good comments and documentation will help you re-acquire the context you need, not completely eliminate it.
Add "Don't add any code comments anywhere" to your system prompt.
If the model doesn't follow this, you want to start using a better model ASAP, because SOTA models for the last year or so, been able to following this without an issue.
Alright, what you'd put instead of "Don't add any code comments anywhere"?
I agree with the general guidance, but it's a general one and not applicable for everything. Some things cannot be expressed in a "do" way rather than "don't".
I don't think changing "don't" to "do not" is what the paper authors had in mind
I use things like "Your code should be self-documenting, so as to require as few comments as possible. Add comments to explain "why" or give important context not apparent from the code itself, but keep them to necessary comments only"
But that's a much laxer rule. I don't think you can truly express "no comments, ever" without a "don't" rule.
Your proposal is still a "Prohibition-type constraint" that the paper you linked earlier say "isn't good".
Some of these constraints we want simply aren't possible without adding "do not" somewhere in the line, even if you prefix/suffix it with other stuff, as you noticed yourself :)
If you don’t trust the code to write a decent comment, why trust it write good code?
Of course, ensuring compilation or other checks can verify some code, which it can’t do for comments. But comments still serve the same purpose as human comments.
If I understand correctly, it’s not that the LLM can’t write a good comment, it’s that you want to be able to interpret and understand the generated code without comments - and in that process end up writing comments yourself.
> Then this one really is a pattern that creates a lot of churn:
> - Add a small, to the point, comment...
As if you even need to tell Claude to add comments. Over the past few weeks I've noticed Claude over-commenting everything. Massive PRs where you realise that fully half or more of the lines that have changed are comments.
It's no good at all: it just pollutes the context, causes token churn, ablates quality, and makes getting to a high quality outcome considerably slower and more expensive.
I get that sometimes knowing why a thing is the way it is can be useful and valuable, but this is what commit comments are for in my mind.
I've had to tell Claude to stop commenting code because the behaviour has become so problematic.
> The what _is_ the code.
Exactly.
If I don't know what the code does because it's arcane and not commented I can simply ask the LLM to explain it to me. I don't need an essay in comment form.
Right. I've really struggling to get AI to stop explaining the what. It seems to add it to the commits, PRs, code, wherever it feels like. I've put in multiple places to not write the "what", but the "why", and in multiple ways, but it still does it in one or other place.
The best way I've found to solve this is using LLM as CI - use a small cheap model to inspect the diff and look for those kinds of comments. Prompt left to the observer but using `claude -p` / `codex exec` gets you a lot cleaner output usually, and makes robots fight robots instead of you constantly having to reprompt and it ignoring you.
Hmm, I guess everyone here is using Claude? I find that Sol is much more restrained, to the point where I have a prompt to tell it to add short comments for things that are not obvious. Really, I find the verbosity problem to be worse in tests. I regularly prompt my review agent to remove tests used only for scaffolding to write the code in the first place. The agent is in a way following strict TDD, which reminds me why I don’t like TDD, even though some of the generated tests can be useful.
Sol adds comments when I use it, but it's not nearly as verbose as Opus 5/Fable.
For the latter I'll often include an example of a comment it wrote, along with my own rephrasing, and tell it that "future readers will understand code context; good naming is the best documentation". This works alright if I include in the actual prompt, but annoyingly it often doesn't in CLAUDE.md or memory.
We trained agents on millions of pages of documentation telling them to write good comments and good code and then we tell them never to write any comments.
It’s almost the “we built a robot who loves to play Sonatas and gave it no hands” type of thing.
What models are you using? I've had "Never add any new code comments" in my system prompts for like a year at this point, seems every model above 14B picks this up just fine. Are you using llama 7b or similar for testing this?
I added to the memory, system prompts, and the prompt itself and every soa model still litters code with the most inane useless crap. I will then get code to review from a coworker using fable/opus. It has more lines of comments then code.
Maybe I am some god tier code reader (i am not) but i dont think i have ever found a comment in code to be useful in my day job. That isnt true, i once came across
// submit to the dark lord
Above the function that sent a payment to PayPal for processing. It made me laugh so I let it be.
I would rather bet that people don’t know that their problem has been solved for ages. Either they don’t know about the tools or can’t make the leap to think of using something like awk or sed to quickly script out their use cases. Or even quickly draft up a quick function/plugin in something like vim, emacs, sublime,…
In “The Pragmatic Programmer”, the power of unix tools and editor fluency is well argued. There are plenty of other books like “Unix Power Tools”, “Small, Sharp Software Tools”,…
Yeah, I've actually found in my own testing and usage of LLMs that this is where I get a lot of benefit. I already have fd, ripgrep, etc. installed and know how to use them, but it's not hard to tell the LLM to do it and it often finds things just as well. Or even better.
It's especially handy on modern style code where things get broken up across a multitude of files based on convention.
This! I also saw lot of potential work that could be done by linting tools. Remember to always prefer mechanical guards than agent instructions, as they cannot ignore them.
The cost of custom linters has like any other code dropped through the floor. I'm sprinkling all kinds of linters over my latest projects. It seems some people are still sleeping on this, expecting great code from the agents.
They're fast and deterministic and I run them in git pre-commit.
Yeah, I'm not sure what people are thinking. I keep reading stuff from folks like "I wish models had more common sense" and "I wish they wrote better code", not realizing this is 100% in your own hands, always been. There is no such thing as "clean code" that every programmer agrees on, you have your own subjective opinion and "good" taste about the code, instruct the models to follow it! And automate it while you're on it.
Incidentally, I'm from the opposite school and consider every “if” followed by a braced block a smell.
If a conditional body needs a block, it's doing enough to deserve a name, so I promote it to a single named call, à la "Extract till you drop".
Another phrase for this is "functional decomposition", which usually is a good thing.
Better yet is to identify conditional execution paths as early as possible in order to obviate conditionals in the call tree. For example, identifying a "create a new something" verses an "update an existing something" based on the workflow initially invoked greatly simplifies service and/or persistent store logic.
Labs have now long understood that ASCII drawing is a core skill needed for coding agents. However, I would not trust them understanding what an existing drawing means, unless it has generated itself.
Linters and static analysis -> setup as hooks in your harness. Don’t rely on CLAUDE.md because it’ll ignore it a lot.
> ASCII drawings in code
Please don’t this is super obnoxious. Make proper diagrams and kee them in knowledge base. Link out to them if you need to and let the agent fetch them via MCP or API or whatever if it wants them.
Even the why sometimes shouldn’t be a comment, unless it’s very immediate to the code itself. What’s often more necessary is a high level overview of the design of the solution, because that’s what drives the design of the code and link disparate section. Especially the glossary , which you let you understand the name of the symbols (variables, struct. functions,…) used in the code.
It’s like learning the culture associated to a foreign language instead of trying to translate each single word with a dictionary.
One thing I don’t get with a lot of these agents.md and other skills are… why not throw as much mechanical checks and other stuff at the repo to constrain as you want instead of asking a non-deterministic agent (squishy or non-squishy) to maintain it.
With the mechanical routes, we get checks, failures, and so much more. A bit wild to me.
Make an agent operate within defined constraints and yell at it when it doesn’t.
Since we are sharing our AGENTS.md, I thought I'd share my own, because most of the time, this is pretty much all you need for LLMs to write good code, everything else can be added per project:
----
*Convergence rule*
Every substantial task must end in exactly one of three states:
A. Success
The intended capability works in the real path and the real motivating case materially improves.
B. Meaningful progression
The capability is not complete, but one genuine blocker is removed and the next blocker is isolated with evidence.
C. Honest stop
Further work would require overbroad scope expansion, excessive debt, brittle patching, or tangled logic. Stop and report the reason with concrete evidence.
Do not continue producing patches once the work stops converging.
Do not confuse activity with progress. A failed attempt is only acceptable if it leaves behind a narrower problem, stronger evidence, or a justified stop.
Any partial work must leave the codebase in a cleaner, more legible, and more diagnosable state than before.
----
A lot of the article's AGENTS.md just feel like telling the LLM agents either something they already know (for example, most of the time they know to use exhaustive switch/match statements instead of "arrow anti-pattern") or seems actively harmful ("keep function names short" seems arbitrary and may cause the LLMs to write weird abbreviations for functions that are harder to read and review.
"Genuine blocker" is mostly there because otherwise LLMs may consider the smallest thing that they couldn't immediately figure out to be blockers and stop without implementing anything. The rule is there to tell the LLM if they can figure out how to resolve the blocker by themselves, they don't have to ask me to help resolve the blocker.
Not very often, but when it happens, usually it's time to sit down and brainstorm architecture with the LLM to figure out how to proceed next instead of looping blindly.
the author referenced this paper which I thought was really interesting https://arxiv.org/abs/2307.03172 - "Lost in the Middle" - are there any other papers a bit like this that summarise lessons to do with applying llms
I'm not necessarily looking for the latest and greatest - more papers that those in the community have coalesced around providing nice ways of summarising problems or as a good example of a specific area.
Great stuff. AGENTS.md is not the ideal place for most of it though. Most of what is shown in this article can go in CODING_STANDARDS.md. The skills that I use find this document when it is needed (writing and reviewing code) so it doesn't pollute context when code is being read.
I also have sub-agent reviews (both of a planning phase and the produced code) that would catch some of these problems and demand revisions. [1]
> - If the prompt indicates that a bug is being fixed, don't write the fix right away. First write the test. Observe it failing. Then write the fix. And observe the test passing.
I always use /tdd [2]. Occasionally it results in some silly tests, but it produces much lower defect code. Its not just for bugs.
This is a problem that people mostly have to solve themselves. Like, I've been working with Claude for almost a year now and I have never once seen it write "Arrow Anti-Pattern" code. That, and much of the rest, would be fluff in my projects. Agent instructions are best learned from experience project-by-project.
Yes, the interesting part about seeing other people's agent.md files, is getting to see what issues they have with working with agents. Seems different people run into very different issues, which probably is caused by how differently we work. So a the file probably should be personalised.
Do you give the model access to the ASD-STE100 spec for reference/review or are you just assuming that enough of it is baked into the model for it to mostly adhere to it?
I would describe this as 13 code writing rules (interpreted to be at least 16 - Starting with reduce code indentation) plus a commit message instruction set which I chose to ignore - because it's style-specific and not interesting to me.
8 or 9 of these rules are not necessary. Basic CS is not something I have needed to ask agents, I use, to follow. eg Explaining that you need explicit interfaces is not a necessary instruction, nor is leveraging early return.
Unclear instructions are of limited utility. What "Let the reader of the code breathe" or "reduce code indentation" means is subjective and will rarely be effective. Maybe the training for the language being used has gaps, which others do not. If you want to measure, ask it to output a string when it applies a rule. You'll figure out what works, what doesn't and how often, quickly.
There's 3 or 4 style choices included.
The rest are not something I would use, but we all get burned by different things so I get it.
One thing I've figured out of that qwen3.6 35b refuses to use 2 space indentation for python code, although it claims to be doing it. I know for a fact it is an easy thing to do.
"- Don't touch blocks of code unrelated to the feature you implement. e.g. Don't add comments to a block of code if you did not create it or modify it. As much as possible try to minimize the number of changed lines when implementing a feature."
The feature where you ask the LLM to fix one thing and it fixes three things.
> As much as possible try to minimize the number of changed lines when implementing a feature
Great way to get LLMs to start making an endless profusion of methods instead of adding parameters to or switching to a richer return type from an existing method, in my experience.
I’m tired of seeing “get_total_rounded_up” + “get_total_float” bloat when a few changes to unrelated code to round floats to ints would keep the method API surface small.
I find too often that models do the opposite - they'll pile small targeted band-aids on code blocks based on new requirements, etc, when having them analyze whether a changed (broader) design would result in a far better overall design?
An earlier version of Gemini used to do this a lot to me back when I used it for some light tinkering around on my projects. "Oh by the way, I fixed a misspelling in a comment file completely unrelated to the feature you asked for, so I fixed that as well, shall I commit everything now?" GAHHH. NO.
These days I have an instruction in my default AGENTS.md to bring issues unrelated to the prompt to my attention when found, but never to just automatically fix them.
Anything that goes into the context window has that going for it. That's a huge part of why Claude's gone absolutely bonkers with genuine, brutal honesty. The system prompt's absolutely stuffed full of those keywords, so now every single output is tainted with that right from the start.
What's the point of agents.md if you just use an LLM on a codebase?
Just say, complete this bit like how the rest is...
Even then they aren't great at it. Idk, the best case use for LLMs are extremely specific requests, for example "write an evaluator for this byte code and if you can't ask for clarification"
The ultimate specification language is code anyways so you might as well stick a to-do, a comment describing the semantics of the function and say "okay codex fill the to-do"
There is an annoying phenomenon with LLMs called "context dilution" or "attention dilution" that was outlined in the Lost in the Middle paper. As the context grows, a model starts paying less attention to instructions in the middle of the context in favor of what is at the beginning and the end.
So then what happens if the agents.md file is colossal and precise and all the required execution instructions are buried in the middle? Do the agents then just fail to execute?
Yeah. That's why people these days avoid long descriptions and instead keep things as short as possible. in my experience, it seems like LLM can't recognize (or less attention value) unless it's structured in a deductive or inductive way
"When writing something intended for human consumption, (comment, commit message, reply to prompt) use as few words as possible. Pick every word meticulously to reduce the volume to a strict minimum. Be down to the point. Less is more."
The irony in this first paragraph using many words and many ways to convey the same message about succintness. But this file is not for human consumtion so different rules should (still?) apply.
I find myself doing this in prompts, I guess it is a way of adding more weight to parts of the context we consider need emphasis, and shows a lack of trust in the llm's abilities to get the message if it is mentioned once.
Would there be any disadvantage to simply having all of the code style stuff in a CONTRIBUTING.md file and the "how to talk to me" stuff in an AGENTS.md file, since code style stuff also applies to human contributions? Similarly, isn't the "how to talk to me" stuff personal and therefore not something that belongs in a repo?
From FAB's AGENT.MD:
> - Avoid magic numbers and strings by extracting recurring or meaningful values into descriptive constants (const) or enums.
---
I've been seeing the same thing with models like GPT5.6 and Opus4.8 in GH Cop CLI. They still introduce magic numbers, and in Scala they often put an entire 10-line Spark expression inside an if condition instead of extracting it into a meaningfully named value to keep "if" readable. I wonder when common sense instructions will be baked into the models.
I'm also doing something similar, as for over-explaining state and momentary decisions, I have yet to find good wording for it too. I had a bit of success by running a reviewer at the end to look over comments and docstrings and judge if it is "evergreen", but feels a bit like a rain-dance.
But in summary: the more bloated your AGENTS.md is, the worse the context consumption gets. The best approach I use is telling the agent to first think about what it needs to do, then choose which rules apply. I got 100% consistency across every area of my projects.
Im against restricting anything related to code length this goes for function names and length, file length ect.
I rather the dots be as close as possible than trust the agent connects the dots.
I dont care if the file is 5000 lines I rather the agent reads one file and get all the context than trust it will read all the need files.
I see so many review skills that puts hard limits on these thing and it just bad.
The function name shouldn't be limited they should be as clear as possible and if for some reason it over 30 chars so be it.
I want to read the function name and the logic and it match exactly. I don't want the agent being lazy because of some limit I set.
In fact I force my agents to write long functions because I specifically tell it not to break out repeated code that doesn't actually deserve a function.
A check on a function input doesn't need to be a function. A auth guard doesn't need to be it own function.
types.go types.ts absolutely the worst file to see in any code base. Put the type next to the code that uses it.
The comments problem is a big one and no matter how hard I try the LLM will always write them verbosely. I find that including the rules in each prompt is fare more effective than once in the context
Hm. I'm curious if this actually helps LLMs iterate on the code, or if it's human nitpicking over things that the author won't really look at? How would you measure?
Personally, I tend to do 3 passes, where I ask the agent to write, and self review; that's been enough to get things functional enough that I don't need to read the code.
the author's correct about using agents but tbh I thought this was widely understood to be the best practice. in fact, agents.md for all your repos and a seperate [claude|gemini|...].md for each repo.
> Explicitly ask the harness to reload agent.md. "Reload agent.md" is enough when I see code quality dropping.
Having the LLM re-read the file is really silly and a common bug in harnesses. Even sillier is when the harness allows a file to be compressed away during summarisation. The harness should compose the context so this doesn't happen. Files should be "added" (by LLM or human) and then always be injected into context the same way forever. "Reload this file" is not something you should ever have to type.
I've read a few of these over the years, and none of them seem to be useful. I have three sentences in my custom instructions, and those are basically all useless, too.
Even my second one, "Avoid decorative or section-header comments. Never use `----` or `====` as comment separators. Comments should explain only non-obvious behavior, rationale, constraints, or implementation details." seems to be ignored by models regularly, so I don't see the point.
But this is in my private harness. Perhaps other harnesses have better instruction following. My custom instructions are prepended to my first user message, not set as a system message.
I honestly think some of this stuff is wrong and constrains the models unnecessarily. Sometimes you're better off with simple adjectives and philosophical directions than getting so specific.
The problem is that linting AGENTS.md is risky. Everything added there was in response to mistakes. If I remove some instruction I run the risk of repeating the mistake.
The most powerful change I have run into is: "Positive phrasing" as a default, prefer to tell the model what they should do, and why. Not a prohibition on a behavior.
When you say "don't do x" you are just pre-seeding the model with "x" and the prohibition mitigates that some, but not as much as never having put "x" in the context in the first place.
"Do Y, for these reasons" can be shaped to achieve what you mean by "Don't do X"
"Don't do X" leads to "Wait, I need to make sure I didn't X" and "Let's look up X to make sure I don't do that." And each time the odds of X happening keeps going up, not down.
These days that sounds like a really good idea. 6 months ago, AGENTS.md would have contained a lot of instructions that would have been embarrassing to write out for a human audience.
No, why should I have to remember to @ in every prompt? Or ask contributors to remember. It just makes it easier to make human mistakes. I have better things to do than micromanagement. There is huge value in auto-included context.
The GP wrote @ it from the agents.md file, not from the prompt. Their point was that instead of writing "how to contribute" instructions for agents, you could explain that in the CONTRIBUTING.md and link it from your agents file, so both humans and agents read it from one place.
this was what i was doing 3-4 months ago. i just have AI write/update my agents.md file now as i find problems. i also have ai keep a set of design documentation that it can update as it goes too. oh and he should try omp+codex/xhigh, he will probably be less annoyed.
My most impactful section has been on voice. It's impact is that I don't go insane, which is pretty high value. (Not putting quote blocks so people can copypasta):
## Voice
Rule #1: No AIisms
Avoid the stock phrases and rhetorical tics that mark AI prose. Say the thing
plainly instead. Be concise and direct.
*Banned phrases* — never use these, or close variants:
- "Honest" or "honestly"
- "Exactly" or "exact, unless referencing a specific quantity or measurement
- "You're absolutely right" / "You're right to push back" / "Great question"
Then this one really is a pattern that creates a lot of churn:
- Add a small, to the point, comment to explain what the block does and why. Use examples when possible. Propose ASCII drawings to explain complete systems.
The what _is_ the code.
reply