yeah MCP contains a provision for some blocktext for instructions. If your MCP has walls of text in those instructions, it will burn more tokens that a terse MCP.
That said, AGENTS.md doesn't seem like a good name, right?, technically, it's an instructions file read by a single agent, not necessarily for agents, so it always struck me as a bit odd
But until the next standardization, keeping just AGENTS.md is the best approach.
In fact, post-LLM quality monitoring is a huge time trap.
I can't write anything for humans using LLMs, because there is so much to change that it ends up like the Ship of Theseus, I would have been faster writing it myself than delegating the task.
I've had some success writing certain explicitly technical documents, with a style guide provided to the LLM that it can match, and then going over it by basically iterating on every sentence in the document and asking "can this sentence be removed?".
Style guides are a big tools people are missing out on. It isn't enough to say "write concisely and technically". Give them a sample. Yea verily these many 5 or 6 years ago, "style transfer" was a big thing that early LLM tech was doing. It's still very good at it.
That said I still tend to cut out at least 25% of the resulting verbiage and adding back another 10% or so more of my own original content even under those circumstances.
Where it has been really helpful is that I tend to want to write in a conversational style that doesn't seem to match most people's expectations of a technical document. The LLMs let me write my way and style-shift it into something closer to what people expect. And LLMs, since let's be honest they're the primary audience nowadays. Which I am not even upset about; I'd rather 5 LLMs read the architecture document than the amortized .2 or so humans I could expect in the same circumstance 5 years ago.
Which also implies, in many cases, I am feeding the AI as much text as I expect to come out, or in some cases, even more, as I am describing context, reasoning, and other things that may impact the writing but are not necessarily repeated in the final text. I factored out a lot of the context into my user-level CLAUDE.md which has helped cut that down a bit.
Having the style guide and proofreading it and linking it is more effort than the lazy slopulists want to put in, so your writing will reach far better than theirs will
This has started to be an issue at my work with code. Everyone started using the thermo-nuclear-code-quality-review skill and, while it does a great job finding consolidation opportunities and architecturally-weak code, it also continues to expand PRs well beyond their scope until you end up revamping far more than you intended...
hey thanks for mentioning that skill, I hadn't heard of it before but I just ran it on a non trivial codebase I've been planning to refactor and it did a great job.
I don't use a lot of "skills" so if there are others you've started using frequently I'd be curious to learn about them.
completely agree. i will admit to taking llm phrases when i am really trying to refine every last detail and it just makes a suggestion that is too good to unsee. but in general it doesn’t feel like it helps directly with the “word choosing” part of the writing task at all (if you are someone who cares about word choice), which is… definitely a pretty big part of the job, lol.
it is great at analyzing the argument, finding inconsistencies, helping you think through what parts should be cut, helping you refine examples or fix the occasional “how do i get this phrase to work correctly in this transition?” kinds of stuff. but anything where LLMs are the prima materia… that stuff literally only makes sense _to me_. which makes sense, because it is written _for_ me, no matter what instructions i actually give it, bc of memories and a million other things. and i say this as a complete maximalist wrt. trying to use llms for absolutely every last thing they possibly can be used for, just to see what it’s like.
i guess i would say that it does very, very little to make the writing process meaningfully faster; it _can_ do _plenty_ to help make your output better though, which is definitely something— just isn’t the thing most people are looking for.
You have to assume that in many situations a human is asking an LLM about the content you've provided, even if they or another human reads some or all of it themselves too. So it's valuable to proactively ask your own LLM in the same way, so you can see what the other person will get out of it. And tune the verbiage until the LLM tells them what you want it to :)
Sorry if this comes off as critical, but are you seriously doing that? I can't imagine optimizing my writing for people choosing to mangle it through a LLM rather than for the people reading it as-is.
Also the word "verbiage" ticks me off as a LLM-ism on the level of "delve" or "smoking gun". Are its users aware that in addition to its usage as somewhat of a synonym of "idiolect", it more commonly means "a profusion of words usually of little or obscure content", i.e. slop?
This is the real crime of llms on writing. They’ve taken perfectly reasonable verbal ticks (the one I miss being able to use is scaffolding) and caused the audience to rebel against them en masse. Verbiage is a great word. I refuse to give it up. It’s load bearing even.
I wrote a contract proposal a few months ago and I absolutely did a few rounds with an LLM to get the result I wanted. Anyone applying for a job should be doing this with their cover letters. You know that HR has a pipeline set up with a prompt like "here's the job description, here's the candidate's info, rate the candidate on a scale from 1-10" so you can do the same and learn how to boost your own signal. I have zero qualms about doing this.
As for "verbiage", I guess that's too bad, I'm going to continue using it anyway. Interesting factoid about the word though. I will say that "utilize" is one of my pet peeves (99% of the time the word you actually want is "use") and I've excised it from my spoken vocabulary, but I like(d) how verbiage rhymes with foliage so I'm more reticent to give that one up.
Fair enough, cover letters and contract proposals are more marketing than actual information, so that's fair game I suppose. Good trick. I wouldn't use it in something meant to transmit actual information though, like a manual or something.
> Anyone applying for a job should be doing this with their cover letters
Sadly people lost the skill to write a standard form concise 3-paragraphs-of-3-sentences cover letter even before keyword-based ATS initiated the beginning of the end. It wasn’t supposed to be a summary of a CV and an essay about your lifelong passion for automated ad-bidding quality checks or some other shit employers delude themselves into believing someone might organically be really into.
I agree and disagree. Yes, it turns into the Ship of Theseus, but I feel it helps jump start things. This was actually the way some writers work I.e. just write fast expecting garbage. Keep molding it like clay until you get what you want.
So much of my time right now is completely rebuilding tutorials that the devs had generated because except for steps like open and save its just wrong and it doesn't work and customers are not going to let that slide.
Here there's no skill involved either. It's not that you have some time to catch as many ducks as possible and if you're good at it you catch more. It's that the 10 € give you the right to catch e.g. 3 ducks, so you spend the time that you need to catch 3 ducks (typically for a 7-year-old like 20 seconds, smaller kids take some more but still not long) and then you look at the score below the ducks, which is totally random. So yeah, quite similar to the claw machine although at least they don't have the farce that you could win an iPhone (Actually you can buy more ducks to get higher scores, so the prizes need to be linear. I have seen parents spend like €50 for their kid to get a toy that could be at most €10).
reply