How long can your skills be before your agent forgets what you told it? — Laurie Voss, Arize AI
https://www.youtube.com/watch?v=XzJD1bvXKjs
AI Engineer · 22 min · 290 sentences · sentence-level transcript from YouTube's caption track; each line begins at the time shown

[00:01] [music]
[00:12] All right.
[00:13] Hello everybody.
[00:15] Thank you for coming to this delightfully nerdy talk.
[00:19] Uh this talk has a really long title.
[00:21] Uh so let me give you the short version up front.
[00:23] You write skills files uh and stuff them full of instructions.
[00:27] At some point the model stops keeping track of all of them.
[00:31] The question is where is that point?
[00:33] At what point have you put too many instructions in your skills files?
[00:36] Uh and the answer has changed a lot in the last year.
[00:40] I'm Lori.
[00:41] I'm head of developer relations at Arise AI.
[00:43] Uh in a former life, I co-founded npm Inc. So some of you may know me from the days of JavaScript.
[00:48] These days I spend a lot of time thinking about AI and how to test it.
[00:53] Uh a few months ago I was at AI engineer in Miami which was a good conference.
[00:57] Uh and I was watching a talk by Dexter Horthy.
[00:59] Uh it was a good talk.
[01:00] It was not about this topic at all.
[01:02] Uh but while he was giving that talk he mentioned as an aside uh that an agent can follow up to about 200 instructions uh before it starts forgetting those instructions.
[01:13] Uh and he then he moved on in his talk and it was entirely an aside.
[01:17] Uh and he he mentioned that that figure is from 2025 so things might be better now.
[01:22] Um, and I stopped listening for a second because I was like, 200 instructions.
[01:27] Uh, is not very many instructions at all.
[01:30] Right?
[01:31] A decent skills file blows past 200 instructions almost immediately.
[01:35] Um, if the user says X, do Y, always include a section on Z, never use the phrase W, every one of those is a separate instruction.
[01:42] Uh, and if the model quietly stops tracking them after 200, that's a really hard ceiling on the complexity of what you can build.
[01:49] Uh, so I wanted to know where he got that number first.
[01:52] Uh, and I wanted to know if it was true.
[01:55] Uh, so you know the feeling that I'm talking about.
[01:57] You write this big beautiful skills file, pages of rules, edge cases, tone, formatting, you hand it to the agent, it does the thing.
[02:03] Uh, and you look at the output and go, did it actually pay attention?
[02:07] Did it actually follow all of these rules or did it just sort of, you know, do what it felt like and sort of give me a close simulacum of what I was expecting?
[02:16] Um, you can't really tell or can you?
[02:18] More on that later.
[02:19] Um, and so you live with this lowgrade anxiety every time you hit run.
[02:24] Um, and that feeling is what this research is about and what we're trying to find out if we can avoid.
[02:30] So here's my promise for your next 18 minutes.
[02:33] Uh, I'm going to show you where that 200 number came from, whether it's still true, and what the real number is today.
[02:39] Uh, because it moved by an order of magnitude.
[02:41] Uh, and then we're going to talk about what that means for you to take away.
[02:46] uh how long your skills and prompts can actually be and what that should what changes you should make to your workflow as a result.
[02:54] So the 200 number isn't folklore.
[02:56] Uh it comes from a real benchmark called IFScale uh from a paper uh by this guy whose name I'm going to mess up Jeroslowitch uh and co-authors last year.
[03:06] And the test is beautifully simple.
[03:09] Uh here's how if scale works.
[03:12] If you ask the model to write a business report uh and you give it a list of specific words that it has to include exactly in the report, include the exact word customer, include the exact word revenue, and so on for as many words as you want.
[03:24] Each of those is an instruction that it has to follow.
[03:27] Uh and then you count how many of those exact words showed up.
[03:31] Um so because the test is so simple, you only have to keep two numbers in your head.
[03:36] One is density, which we call n. That is how many rules we're talking about at once.
[03:41] And the second is accuracy, which is the percentage of those rules uh that it was able to actually follow.
[03:47] Uh now you might say uh that including random words in a report is not the same as following real instructions and fair enough and we're going to talk about that.
[03:55] Um but the keywords are a proxy.
[03:57] Uh include the word revenue is the same shape of task as include a section on pricing, right?
[04:03] Or never use this phrase.
[04:04] It is a discrete named constraint that you've told the agent that it has to follow.
[04:08] Um, if a model can't track 200 words in one prompt, it's definitely going to struggle with 200 more complicated instructions.
[04:16] Uh, so uh, if anything, it's going to do worse.
[04:20] So, this number is a ceiling.
[04:21] Uh, this number is as high as you can go.
[04:24] If you give it more complicated instructions, the number is probably going to get lower.
[04:27] And 200 is a really low ceiling.
[04:30] Um so before chasing new models you have to do good science which means that you have to replicate the uh old result and make sure uh that the 200 ceiling is real.
[04:39] So I reran the original benchmark.
[04:42] Um the original paper tested a whole batch of models uh and models uh live and die really fast.
[04:48] So uh by the time I got around to doing this testing only three of the models in the original set of 10 models that they used were still available via any kind of API.
[04:56] Uh so they were GPT 4.1, Claude Sonnet 4, and Gemini 2.5 Pro.
[05:01] Those were models that were available 12 months ago that are still available now.
[05:05] Um and that is why we tested those three because they were what was left.
[05:08] Um and since I first published this research a couple of weeks ago, uh one of those three models has been retired.
[05:14] So this was the last possible time that I could have run this test.
[05:18] Um so of that lineup, we're already down to two.
[05:20] So don't get attached to your models.
[05:22] Um here is the results that we got replicating the original if scale finding.
[05:27] Uh that is accuracy on the vertical axis.
[05:29] So it starts at 100% and begins to fall off.
[05:32] Uh and then the number of rules uh going up along the bottom on log scale.
[05:36] So every time it gets halfway across it has doubled uh the number of rules that it's dealing with.
[05:41] Um so by 500 rules you're losing 30 40 50% of them.
[05:46] Uh our curves matched the results in the original paper within the noise boundary.
[05:50] So the finding was real.
[05:52] uh a year ago somewhere around 200 to 300 rules frontier models started falling apart.
[05:57] That is a really low ceiling.
[06:00] Uh so that is our baseline and now comes the fun part where we took the exact same test and pointed it at the current frontier or rather what the current frontier was when I ran this test.
[06:10] So I ran GPT 5.5, Claude Opus 4.7 because 4.8 came out a week after I ran this test.
[06:16] Uh Gemini 3.1 Pro and Deepseek V4 Pro.
[06:20] So, I gave them the same prompt, the same words, the same everything.
[06:23] And I immediately ran into a problem, which is that they aced it.
[06:27] They all scored 100% immediately on this test.
[06:30] Absolutely no bugs.
[06:32] Uh, so we'd built a test to find the ceiling and the models had walked straight through the ceiling without noticing that the ceiling was there.
[06:39] Um, and that was a problem because the benchmark was written to top out at 500 words.
[06:43] So, I had to change the benchmark in order to be able to find the new ceiling.
[06:47] So, I moved the goalposts.
[06:48] I gave it more words to include.
[06:50] I doubled uh it from 500 to a,000.
[06:52] I doubled it again from a,000 to 2,000.
[06:54] And I kept doing that until I hit a 10,000word vocabulary.
[06:58] And that is where I began to find the ceiling of what models can do these days.
[07:03] Um, so let me put up the this is the money slide.
[07:06] This is the results.
[07:07] Remember log scale on the on the uh x-axis there.
[07:12] So it's going from 500 to 1,000 to 5,000 to 10,000.
[07:15] Uh so it looks like that scale is falling off of a cliff and it's actually happening over like a thousand numbers.
[07:20] Um but uh look how far to the right these new curves get before they bend.
[07:26] A year ago they were falling over at 200 to 300 instructions and now depending on the model the boundary is closer to 2,000.
[07:32] And for the best of them it is up to 5,000 instructions before they begin to fall off a cliff.
[07:38] So in about 12 months frontier models got close to 10 times better at following instructions simultaneously.
[07:44] That is the headline fighting and there is a lot of nuance that we need to get into.
[07:49] Um the capacity to track 2,000 named constraints in a single prompt is there.
[07:54] Um and that's really interesting because I think uh I don't know if everybody else feels this way but like it feel it felt to me like the the jump from you know GPT 5.1 to GPT 5.5 was kind of incremental, right?
[08:07] It didn't feel like we'd got 10 times better.
[08:10] But this is a test that really matters to uh a very practical thing like how long can my skills file be?
[08:17] Uh and in the course of a year we got 10 times better.
[08:20] Uh and that thing that gets me is that this benchmark is barely a a year old.
[08:25] A year later 500 is a rounding error.
[08:27] Uh and this keeps moving under my feet.
[08:29] I tested 4.7 uh opus 4.7. Opus 4.8 is even better.
[08:33] Um so this chart is a little out of date already which is kind of the whole point.
[08:37] If you set your engineering assumptions about how skills files should work, about how prompt how long your prompt can be, and you did that more than about six months ago, you are incorrect now, and you should probably be re-engineering
[08:49] how you do stuff.
[08:51] Uh, but there is more to this story uh because the way that the a models failed uh changed dramatically.
[09:00] Uh, and the way that they failed is very important.
[09:03] This part was a completely unexpected finding when I started running the experiment.
[09:08] Uh, and it totally messed up my test to start with because uh, the old failure mode was boring.
[09:13] They would just forget instructions and I could measure how many instructions they had remembered or forgotten.
[09:17] Uh, but the new ones fall apart in their own weird extremely onbrand way.
[09:23] Uh, so let me introduce you to how these four models fail.
[09:26] Uh, Deepseek 4 is a traditional model.
[09:29] It just forgets things.
[09:30] It doesn't have any drama.
[09:32] um it starts forgetting instructions around 750 rules and by 2000 it's dropping nearly half of them.
[09:38] Uh so it just forgets which frankly is the failure mode that I trust most because it's predictable.
[09:43] It's very easy to measure.
[09:44] Uh and the other models were not nearly as cooperative.
[09:48] Uh Opus 4.7 uh would decide repeatedly that the test was dangerous.
[09:54] Uh and what it would do is it would refuse at the API level to complete the test.
[09:59] I didn't know that there was an API response that you could get from Claude where it was like, "No, I could do this, but I'm not going to."
[10:06] Uh, but that's absolutely an API level response that Claude supports because they care so much about safety.
[10:12] Uh, and I started getting those all of the time.
[10:15] Uh, and the reason that was happening is because Claude has a very sensitive safety classifier.
[10:19] Uh, and if you put in certain combinations of words like say anthrax and cyanide, it decides that the whole request is dangerous and it bails out.
[10:27] Uh, and if you remember what my test does, my test is throwing uh 5 to 10,000 random words into uh into an instruction file.
[10:35] And so my my randomly selected words contained all sorts of things that looked dangerous in combination to the safety filter.
[10:41] And so it kept bailing saying that I was asking it to, you know, make a bomb or something.
[10:45] Um, so, uh, we had to for to get Claude to cooperate, I had to take all of my words and run them through OpenAI safety filter and filter out all of the naughty looking words so that it could get to anywhere.
[10:59] Once I given it that, Claude did really well.
[11:02] Uh so but the failure mode is that Claude is more likely to decide what you're doing is dangerous very early on uh at you know even two or 300 instructions if what you're doing
[11:12] uh is you know contains anything to do with medical advice because medical things often are dual purpose.
[11:17] They can be dangerous.
[11:18] They can be safe.
[11:19] Um so uh the third failure mode was Gemini 3.1 Pro.
[11:24] Gemini is rock solid all the way out uh to 5,000 instructions.
[11:28] It does extremely well.
[11:30] um genuinely one of the best on the chart.
[11:32] Uh and then past that it gets weird.
[11:34] Um it doesn't forget the instructions, it gets overwhelmed by the instructions.
[11:39] What it tries to do is it uh it uses thinking tokens to make sure that it is following all of the instructions at once.
[11:46] And when the number of instructions gets really high, it uses all of its thinking tokens.
[11:50] It uses its entire token budget thinking.
[11:53] And then it doesn't give any output.
[11:55] It's it gets to like nine, you know, if you've given it 10,000 tokens worth, it'll get 9,500 tokens worth of thinking and then give you a 500word response which doesn't contain any of the tokens.
[12:06] Uh so it thinks itself into a corner and runs out of room to actually answer, which is very expensive, uh and totally unhelpful, which is kind of on brand, isn't it?
[12:16] Um
[12:17] [snorts]
[12:18] uh which, you know, I would never say that out loud.
[12:21] Uh, and finally comes the winner, which is uh, GPT 5.5. GPT 5.5 is the best of the lot.
[12:28] 99% accuracy all the way out to 5,000 rules.
[12:31] Um, but if you push it far enough, it is by far the weirdest of the bunch.
[12:35] Uh, because it doesn't refuse outright.
[12:37] It doesn't silently forget.
[12:39] Instead, what it does is it gets frustrated and tells you that the test is stupid.
[12:44] Uh, it starts the report.
[12:46] It gets a few like that's the thing.
[12:48] It doesn't start out just saying no. It starts the report, it starts writing the report, and like 500 words into the report, it's like, "No, this is dumb.
[12:55] I'm not going to do this."
[12:56] And then it politely tells you, "This is dumb.
[12:58] I'm not going to do this anymore."
[13:00] Uh, that is the actual response that it gave me, but that that was like 5,000 words into the into this business report that I told it to generate.
[13:06] Um, so it's not wrong, right?
[13:10] I was asking for a coherent business report that on no particular subject that contains 5,000 random words.
[13:17] You're right, Gemini DPT.
[13:19] this is a a stupid thing to ask for.
[13:21] Um, which is a deeply unreasonable request and GPT called this out on it.
[13:26] Um, but it still counts as a failure in the test because the half-finish report that it gives you is missing most of the keywords and it is also the hardest one to detect because claude bails immediately.
[13:37] Claude says, "No, I'm not going to do this."
[13:39] Uh, Deepseek does its best.
[13:41] Uh, but GPT does what looks like a good job unless you read all the way to the end of the report where it says, "No, actually I'm going to bail because this is stupid."
[13:50] Um, so if you step back and look at the four together, Deep Sea quietly forgets, Claude gets scared and refuses, Gemini overthinks itself into silence, and GPT 5.5 finishes half of the job and tells you that the rest of it is beneath it.
[14:03] Um, and the point was the point isn't which one of these is funniest, although it is genuinely a little funny.
[14:09] Uh the point is that did it follow my instructions no longer has one failure mode.
[14:14] It has four different ways that it can fail and you can't recognize that failure unless you know which model you're dealing with and what its m what its pattern of failure is going to be.
[14:24] Uh so the models get 10 got 10x better.
[14:27] They fail in funny ways.
[14:29] Why should you care when you uh get back to your desk?
[14:32] Because three things have changed to your workflow.
[14:35] The first is that a year ago, the smart move was to keep every skills file very very short.
[14:40] Uh under 200 instructions, then point off to subsklls and a whole like you know byzantine labyrinth of uh additional skills files and subfiles and things like that.
[14:51] Uh and you mo you were compressing your your instructions to fit into a very small available space and you don't need to do that anymore.
[14:58] Your skills files can be very long.
[15:00] Um, number two is that if your use case needs a 100 specific rules or 300, you can just put them all in the prompt.
[15:07] Uh, you don't have to lie awake wondering whether which ones the model silently ignored.
[15:13] Um, and if you've been thinking uh about uh your own lived experience of using models, uh you probably recognize this.
[15:21] you've discovered that you've got less worried about how long your your prompt is going to get uh because the models have genuinely got 10 times better at following your prompts.
[15:30] Um 2,000 named constraints is an entire style guide, right?
[15:35] Like it's it's every brand rule, every legal disclaimer.
[15:38] Uh a year ago, you'd have had to shard that across a dozen specialized agents and hope that your specialized agents are hand are are handing off to each each other cleanly.
[15:47] But now you can ignore that.
[15:48] Um but the third thing is the big one.
[15:51] The question used to be can the model even do this?
[15:53] And the answer is now firmly yes.
[15:55] Well reasonably firmly.
[15:57] Uh is it worth the cost is the new question because you can include 10,000 words of of sorry 10,000 different instructions into your prompt.
[16:06] But that is going to be an enormous prompt.
[16:08] It's going to be a very expensive prompt.
[16:09] It's going to be a very slow prompt.
[16:11] So what used to be a hard wall that you would run against has now become a soft trade-off of is it worth me adding all of these extra instructions if it's going to give me more cost and more latency.
[16:21] Uh and now some caveats uh to head off the Q&A.
[16:26] Um first and important first and most important I mentioned this earlier this is a proxy task including random words uh in a in a fake business report um is evidence
[16:36] that long skills file works.
[16:38] It is not the same as proof that a long skills file works.
[16:41] Um, also the models hit the wall at wildly different points anywhere from 750 to 9,000 plus.
[16:47] So you have to pick your model very carefully.
[16:50] Uh, what our test doesn't do is measure whether the model reasoned clearly over a giant prompt.
[16:58] So uh, the good news is since I did my research several weeks ago, uh, a whole bunch of people have piled in on this.
[17:04] Um and now there's good research uh actual scientists have got involved and done uh Chroma's has done context rot work uh across 18 models showing that accuracy on long inputs
[17:17] can fall 30 to 50% well before you hit the context window limit.
[17:21] Uh and the weird part of their finding was that uh coherent well ststructured text is more likely to hit that failure mode uh than if you just put your instructions into a random order and shuffle them in.
[17:33] Uh, I don't know why that's the case.
[17:36] I'd have to read their report.
[17:37] Um, so the model can track 2,000, 5,000, possibly 10,000 instructions, but it's not necessarily going to uh reason clearly over them.
[17:47] It's not necessarily if those if those instructions conflict, if there is tension between them, it's not necessarily going to get that right.
[17:54] Um, and then there's the other one I mentioned briefly.
[17:57] Uh, collude's refusals are annoying, but they are loud.
[18:01] You get an error, you know it failed.
[18:02] Uh GPT's polite half-finish report is much more dangerous because it looks like a real answer.
[18:07] Uh you have to read the whole thing to notice that it gave up quietly halfway, which means that you can't trust the output.
[18:14] It means you have to read the output every single time to make sure whether or not it's working.
[18:18] Uh so the model will accept your 2,00 rules and it will hand you back something that looks at least to begin with confident and polished but could be bailing out halfway through.
[18:28] Um, so, uh, as an aside, people always ask me, "How much did all this cost me?"
[18:34] It cost me $29 to run all of these queries.
[18:37] 2,37 2,300 calls across seven models, uh, came to $29.
[18:43] Uh, it turns out novel research doesn't cost very much.
[18:46] Um, and this is the part of the talk where I was saying that you have to check this stuff in production because you can't trust that your model isn't going to
[18:55] silently fail.
[18:56] Uh, so you knew I was going to mention evals eventually because I work at Arise and this is where I do that.
[19:01] Um, but there are plenty of plugs for Arise.
[19:03] So I'm just going to say one true thing which is that if you are building a real AI application and you are giving it genuinely tricky tasks, you are going to run into one or more of these failure modes with a frontier model.
[19:14] Uh, and unless it's Claude telling you just to off at the API level, the only way to know that something went wrong is monitoring your outputs with another LLM.
[19:23] That is an eval.
[19:24] And that is what Arise does.
[19:25] And I'll leave it at that.
[19:27] Uh, I already mentioned that there's been new research since we did our own.
[19:31] Here's another important one.
[19:32] A paper landed testing 46 models called revisiting the reliability of language models in instruction falling, which you can bet made my ears perk up after I did that research myself.
[19:41] Uh, and they found something uncomfortable, which is that a model can ace a benchmark like ours and still be wildly unreliable.
[19:47] because if you reword the same instruction in a slightly different way, it can make a radical difference to how well uh it follows those instructions.
[19:57] So the model can follow 2,000 instructions and it can do it really well.
[20:00] But if you put the same instructions, the same 2,000 instructions in a different order, it can suddenly make the model much worse at following those instructions.
[20:09] And how ex how exactly to do that?
[20:12] what is the correct order of instructions to give your model such that it follows them perfectly as opposed to getting confused is still research that is being done.
[20:20] So capacity went up but reliability is still a problem.
[20:24] Um and then this is just a little brag because uh I was happy about it like I'm not a scientist.
[20:30] I did some research and then a whole bunch of other actual scientists piled in uh and did real science on the same question.
[20:36] There's now a whole bunch of benchmarks that have shown up uh to measure this same question.
[20:40] Firebench, CCR bench, Guidebench uh are all trying to measure the same thing.
[20:44] How well models follow a lot of real messy constraints at once.
[20:49] Uh and now the whole field is looking at it.
[20:51] So if you want better science than my, you know, 10,000 random words, uh the real science exists now.
[20:58] Uh so that gets me to where I will leave you.
[21:00] A year ago, the hard part of writing a skill was fitting everything in without the model losing the plot.
[21:05] That was a compression problem, and the compression problem is gone.
[21:09] uh the model will hold your 2,000 instructions just fine.
[21:12] The new hard part is knowing whether it actually did what you said and that is a verification problem.
[21:17] Uh a verification problem doesn't get solved by writing a better prompt.
[21:20] It gets solved by checking the output every time uh the same way that you would test any other code, which is to say an eval.
[21:27] The ceiling moved by 10x in one year.
[21:29] Uh so go back and check the assumptions that you made six months ago about how big your prompts should be, how big your uh instructions can get.
[21:38] uh because they might already be wrong.
[21:39] Boom.
[21:40] Be wrong.
[21:41] So that is the talk.
[21:42] If you want uh all of the code and all of the data, uh it is at this GitHub URL.
[21:47] Uh and this other QR code is uh something marketing made me insert.
[21:52] We are having a World Cup watch party tonight at 5:00 p.m. Uh you can come to our party.
[21:56] That link is to the Luma that will get you into the get into get you into the party.
[22:01] Uh I hope this talk has given you some novel information or at least a couple of laughs.
[22:06] And thank you so much for your time and attention.