There is much missing from this prompt, tool call descriptors is the most obvious. See for yourself using even a year old jailbreak [1]. There’s some great ideas in how they’ve setup other pieces such as cursor rules.
They use different prompts depending on the action you're taking. We provided just a sample because our ultimate goal here is to start A/B testing models, optimizing prompts + models, etc. We provide the code to reproduce our work so you can see other prompts!
The Gist you shared is a good resource too though!
Maybe there is some optimization logic that only appends tool details that are required for the user’s query?
I’m sure they are trying to slash tokens where they can, and removing potentially irrelevant tool descriptors seems like low-hanging fruit to reduce token consumption.
I definitely see different prompts based on what I'm doing in the app. As we mentioned there are different prompts for if you're asking questions, doing Cmd-K edits, working in the shell, etc. I'd also imagine that they customize the prompt by model (unobserved here, but we can also customize per-model using TensorZero and A/B test).
Yes this is one of the techniques apps can use. You vectorize the tool description and then do a lookup based on the users query to select the most relevant tools, this is called pre-computed semantic profiles. You can even hash queries themselves and cache tools that were used and then do similarity lookups by query.
Cursor and other IDE modality solutions are interesting but train sloppy use of context.
From the extracted prompting Cursor is using:
> Each time the USER sends a message, we may automatically attach some information about their current state…edit history in their session so far, linter errors, and more. This information may or may not be relevant to the coding task, it is up for you to decide.
This is the context bloat that limits effectiveness of LLMs in solving very hard problems.
This particular .env example illustrates the low stakes type of problem cursor is great at solving but also lacks the complexity that will keep SWE’s employed.
Instead I suggest folks working with AI start at chat interface and work on editing conversations to keep clean contexts as they explore a truly challenging problem.
This often includes meeting and slack transcripts, internal docs, external content and code.
I had this thought as well and find it a bit surprising. For my own agentic applications, I have found it necessary to carefully curate the context. Instead of including an instruction that we "may automatically attach", only include an instruction WHEN something is attached. Instead of "may or may not be relevant to the coding task, it is up for you to decide"; provide explicit instruction to consider the relevance and what to do when it is relevant and when it is not relevant. When the context is short, it doesn't matter as much, but when there is a difficult problem with long context length, fine tuned instructions make all the difference. Cursor may be keeping instructions more generic to take advantage of cached token pricing, but the phrasing does seem rather sloppy. This is all still relatively new, I'm sure both the models and the prompts will see a lot more change before things settle down.
Yea the proxying/observability is without question the simplest part of this whole problem space. Once you get into the weeds of automating all the eval and prompt optimizing, you realize how irrelevant wireshark actually is in the feedback loop.
But I also like you landed on mitmproxy as well, after starting with tcpdump/wireshark. I recently started building a tiny streaming textual gradient based optimizer (similar to what adalflow is doing) by parsing the mitmproxy outputs in realtime. Having a turnkey solution for this sort of thing will definitely be valuable at least in the near to mid term.
wireshark would work for seeing the requests from the desktop app to Cursor’s servers (which make the actual LLM requests). But if you’re interested in what the actual requests to LLMs look like from Cursor’s servers you have to set something like this up. Plus, this lets us modify the request and A/B test variations!
Sorry, can you explain this a bit more? Either you're putting something between your desktop to the server (in which case Wireshark would work) or you're putting something between Cursor's infrastructure and their LLM provider, in which case, how?
we're doing the latter! Cursor lets you configure the OpenAI base URL so we were able to have Cursor call Ngrok -> Nginx (for auth) -> TensorZero -> LLMs. We explain in detail in the blog post.
I've been curious to see the process for selecting relevant context from a long conversation. has anyone reverse engineered what that looks like? how is the conversion history pruned, and how is the latest state of a file represented?
[1]: https://gist.github.com/lucasmrdt/4215e483257e1d81e44842eddb...
The Gist you shared is a good resource too though!
I’m sure they are trying to slash tokens where they can, and removing potentially irrelevant tool descriptors seems like low-hanging fruit to reduce token consumption.
From the extracted prompting Cursor is using:
> Each time the USER sends a message, we may automatically attach some information about their current state…edit history in their session so far, linter errors, and more. This information may or may not be relevant to the coding task, it is up for you to decide.
This is the context bloat that limits effectiveness of LLMs in solving very hard problems.
This particular .env example illustrates the low stakes type of problem cursor is great at solving but also lacks the complexity that will keep SWE’s employed.
Instead I suggest folks working with AI start at chat interface and work on editing conversations to keep clean contexts as they explore a truly challenging problem.
This often includes meeting and slack transcripts, internal docs, external content and code.
I’ve built a tool for surgical use of code called FileKitty: https://github.com/banagale/FileKitty and more recently slackprep: https://github.com/banagale/slackprep
That let a person be more intentional about what the problem they are trying to solve by only including information relevant to the problem.
(that being said, mitmproxy has gotten pretty good for just looking lately https://docs.mitmproxy.org/stable/concepts/modes/#local-capt... )
But I also like you landed on mitmproxy as well, after starting with tcpdump/wireshark. I recently started building a tiny streaming textual gradient based optimizer (similar to what adalflow is doing) by parsing the mitmproxy outputs in realtime. Having a turnkey solution for this sort of thing will definitely be valuable at least in the near to mid term.
We plan to continue investigating how it works (+ optimize the models and prompts using TensorZero).
I wonder how hard it would be to build a local apply model/surely that would be faster on a macbook