I've seen Claude ignore important parts of skills/agent files multiple times. I was running a clean up SKILL.md on a hundred markdown files, manually in small groups of 5, and about half the time it listened and ran the skill as written. The other half it would start trying to understand the codebase looking for markdown stuff for 2min, for no good reason, before reverting back to what the skill said.
Try this: Keep your CLAUDE.md as simple as possible, disable skills, and request Opus to start a subagent for each of the files and process at most 10 at a time (so you don't get rate limited) and give it the instructions in the skill for whatever processing you're doing to the markdowns as a prompt, see if that helps.
yes we had to tune the claude.md and the skill trigger quite a bit, to get it much better. But to be honest also 4.6 did improve it quite a bit. Did you run into your issues under 4.5 or 4.6?
I have seen numbers claiming tools are only called 59% of the time.
Saw another comment on a different platform where someone floated the idea of dynamically injecting context with hooks in the workflow to make things more deterministic.
insights is straight ego fluffing - it just tells you how brilliant you are and the only actionable insights are the ones hardcoded into the skill that appear for everyone. things like be very specific with the success criteria ahead of time (more than any human could ever possibly be), tell the llm exactly what steps to follow to the letter (instead of doing those steps yourself), use more skills (here's an example you can copy paste that has 2 lines and just tells it to be careful), and a couple of actually neat ideas (like having it use playwright to test changes visually after a UI change)
Some people just can't take a compliment, especially if it's generated. (I'm one of them.) Still, /insight did give useful help, but I wasn't able to target it to specific repo/sessions.
Ohh this is exciting, I kinda overlooked it. I assume there are still a lot of differences, especially for accross teams. But I immediately ran it, when I saw your comment. Actually still running.
> 26% of sessions are abandoned, most within the first 60 seconds
Starting new sessions frequently and using separate new sessions for small tasks is a good practice.
Keeping context clean and focused is a highly effective way to keep the agent on task. Having an up to date AGENTS.md should allow for new sessions to get into simple tasks quickly so you can use single-purpose sessions for small tasks without carrying the baggage of a long past context into them.
this jumped out at me too. What counts as "abandoned"? How do you know the goal was not simply met?
I have longer threads that I don't want to pollute with side quests. I will pull up multiple other chats and ask one or two questions about completely tangential or unrelated things.
I abandon sessions when I ask for something then it spins for a minute, fills up 40% of the context window and comes back with the totally wrong questions and I don't like the approach it took to get there. I don't answer any of the questions and just kill the session and start a new one with a different prompt.
From session analysis, it would be interesting to understand how crucial the documentation, the level of detail in CLAUDE.md, is.
It seems to me that sometimes documentation (that's too long and often out of date) contributes to greater entropy rather than greater efficiency of the model and agent.
It seems to me that sometimes it's better and more effective to remove, clean up, and simplify (both from CLAUDE.md and the code) rather than having everything documented in detail.
Therefore, from session analysis, it would be interesting to identify the relationship between documentation in CLAUDE.md and model efficiency. How often does the developer reject the LLM output in relation to the level of detail in CLAUDE.md?
Its our own sessions, from our team, over the last 3 months. We used them to develop the product and learn about our usage. You are right, they will remain closed. But I am happy to share aggregated information, if you have specific questions about the dataset.
> A local-first desktop and web app for browsing, searching, and analyzing your past AI coding sessions. See what your agents actually did across every project.
Our focus is a little bit more cross team, and in our internal version, we have also some continuous improvement monitoring, which we will probably release as well.
Hey, here is Rafa, another Rudel AI developer. The ultimate goal is to make developers more productive. Suddenly, we had everyone having dozens of sessions per day, producing 10X more code, we were having 10X more activity but not necessarily 10X productivity.
With this data, you can measure if you are spending too many tokens on sessions, how successful sessions are, and what makes them successful. Developers can also share individual sessions where they struggle with their peers and share learnings and avoid errors that others have had.
Big ask to expect people to upload their claude code sessions verbatim to a third party with nothing on site about how it's stored, who has access to it, who they are... etc.
We dont expect anything, we put it out there, and we might be able to build trust as well, but maybe you dont trust us, thats fair. You can still run it yourself. We are happy about everyone trying it out, either hosted or not. We are hosting it, just to make it easier for people that want to try it, but you dont have to. But you have a good point, we should probably put more about this on the website. Thanks.
the skill usage was one of these "I am wondering about...." things, and we just prompted it into the dashboard to undertand it. We have some of these "hunches" where its easier to analyze having sessions from everyone together to understand similarities as well as differences.
And we answered a few of those kinda one off questions this way. Ongoing, we are also using a lot our "learning" tracking, which is not really usable right now, because it integrates with a few of our other things, but we are planning to release it also soon.
Also the single session view sometimes helps to debug a sessions, and then better guide a "learning".
So its a mix of different things, since we have multiple projects, we can even derive how much we are working on each project, and it kinda maps better than our Linear points :)
One potential reason for sessions being abandoned within 60 seconds in my experience is realizing you forgot to set something in the environment: github token missing, tool set for the language not on the path, etc. Claude doesn't provide elegant ways to fix those things in-session so I'll just exit, fix up and start Claude again. It does have the option to continue a previous session but there's typically no point in these "oops I forgot that" cases.
I was about to say they have a self-hosting guide, but I see they use third party services that seem absolutely pointless for such a tiny dataset. For comparison, I have a project that happily analyzes 150 million tokens worth of Claude session data w/some basic caching in plain text files on a $300 mini pc in seconds... If/when I reach billions, I might throw Sqlite into the stack. Maybe once I reach tens of billions, something bigger will be worthwhile.
1. can only partly be answered, because we can only capture the "edits" that are prompted, vs manual ones.
2. for us actually all of them, since we do everything with ai, and invest heavily and continously, to just reduce the amount of iterations we need on it
3. thats a good one, we dont have anything specific for debugging yet, but it might be an interesting class for a type of session.
To clarify, our data set consists solely of Claude Code sessions, specifically those with a human behind them. Rudel AI, in its current form, focuses on "How teams code with AI". We have plans to expland to a larger range of agentic observability use cases.
This is great. How are you "identifying" these stages in the session? Or is it just different slash commands / skills per stage?
If its something generic enough, maybe we can build the analysis into it, so it works for your use case. Otherwise feel free to fork the repo, and add your additional analysis. Let me know if you need help.
I use prompt templates, so in the first version of my analysis script on my own logs I looked for those. However, to make it generic, I switched to using gemini as a classifier. That's what's in the repo.
I usually instruct the agent to use the skills explicitly, e.g. "/writing-tests write the tests for @some-class.cpp"
So the skills are mostly a sort of on-demand AGENTS.md specific to the task.
Another example is I have a `plan-review` skill, so when planning something I add at the end of the prompt something like: "plan the task, .... then launch claude and codex /plan-review agents in parallel and take their findings into account before producing the final plan".
The 4% usage was about our internal team, and we have skills setup. So it is not necessary that they are not built, but rather that they were not used, when we expected them to be used. So we adapted our CLAUDE.md to make claude more eager to use them. Also the 4% usage was on the 4.5 models, 4.6 got much better with invoking skills.
It's crazy how fast I'm able to identify these bots now. You just get an uncanny valley type of feeling immediately reading it. Sure enough you click the profile and it's a brand new account with one or two similar posts in the same style. There's some sort of writing style here that identifies it because I've picked upon it multiple times quickly but it's hard to articulate into words.
Heavy use of /rewind helps with this - it's much better to remove the bad information from the context entirely instead of trying to tell the model "actually, ignore the previous approach and try this instead"
> The 26% abandonment rate, the error cascade patterns in the first 2 minutes — these are behavioural signals, not just performance metrics.
> When Claude Code gets stuck in a loop, tries an unexpected tool chain, or produces inconsistent outputs under adversarial prompts — those aren't just UX failures, they're security surface area.
Twice in one paragraph, not even trying to blend in.
The tool does have a quite detailed view for individual sessions. Which allows you to understand input and output much better, but obviously its still mysterious how the output is generated from that input.
LLMs are far from consistent.
This works in my experience
Saw another comment on a different platform where someone floated the idea of dynamically injecting context with hooks in the workflow to make things more deterministic.
Does this include the files being worked on by the agent in the session, or just the chat transcript?
if you dont trust us with that data though (which i can understand) you can host that thing locally on your machine
Starting new sessions frequently and using separate new sessions for small tasks is a good practice.
Keeping context clean and focused is a highly effective way to keep the agent on task. Having an up to date AGENTS.md should allow for new sessions to get into simple tasks quickly so you can use single-purpose sessions for small tasks without carrying the baggage of a long past context into them.
I have longer threads that I don't want to pollute with side quests. I will pull up multiple other chats and ask one or two questions about completely tangential or unrelated things.
It seems to me that sometimes it's better and more effective to remove, clean up, and simplify (both from CLAUDE.md and the code) rather than having everything documented in detail.
Therefore, from session analysis, it would be interesting to identify the relationship between documentation in CLAUDE.md and model efficiency. How often does the developer reject the LLM output in relation to the level of detail in CLAUDE.md?
I do not see any link or source for the data. I assume it is to remain closed, if it exists.
but i think the prior on 'this team fabricated these findings' is v low
TBH, I am very hesitant to upload my CC logs to a third-party service.
I scrolled through and didn’t see enough to justify installing and running a thing
Thx for the link - sounds great !
With this data, you can measure if you are spending too many tokens on sessions, how successful sessions are, and what makes them successful. Developers can also share individual sessions where they struggle with their peers and share learnings and avoid errors that others have had.
No, thanks
Or you can run your own instance, but we will need to add docs, on how to control the endpoint properly in the CLI.
would love to know your actual day to day use case for what you built
I would say roughly equal amount of sessions between them (very roughly)
Also maybe 40% of coding sessions in large brownfield project. 50% greenfield, and remaining 10% non coding tasks.
What tools do you use to run your analysis?
So the skills are mostly a sort of on-demand AGENTS.md specific to the task.
Another example is I have a `plan-review` skill, so when planning something I add at the end of the prompt something like: "plan the task, .... then launch claude and codex /plan-review agents in parallel and take their findings into account before producing the final plan".
> When Claude Code gets stuck in a loop, tries an unexpected tool chain, or produces inconsistent outputs under adversarial prompts — those aren't just UX failures, they're security surface area.
Twice in one paragraph, not even trying to blend in.
It became very hard to understand what exactly is sent to LLM as input/context and how exactly is the output processed.