4 ms·

I have a daily journal where I put every online post I make. I include anything I send to a LLM on my own time in there. (I have a separate work log on their co

I have a daily journal where I put every online post I make. I include anything I send to a LLM on my own time in there. (I have a separate work log on their computer, though I don't log my work prompts.) Likely I miss a few posts/prompts, but this should have the vast majority. A few caveats: I'm not a heavy LLM user (this is probably what you're getting at) and the following is a low estimate. Often, I'll save the URL only for the first prompt and just put all subsequent prompts under that one URL. Anyhow, running a simple grep command suggests that I have at least 82 prompts saved. In my view, it would be better to organize saved prompts by project. This system was not set up with prompt disclosure in mind, so getting prompts for any particular project would be annoying. The point is more to keep track of what I'm thinking of at a point in time. Right now, I don't think there are tools to properly "share the prompts" at the scale you mentioned in your other comment, but I think we will have those tools in the future. This is a real and tractable problem. > Whats the worth of digging through that? What do you learn? How would you know that I shared all of them? The same questions could be asked for the source code of any large scale project. The answers to the first two are going to depend on the project. I've learned quite a bit from looking at source code, personally, and I'm sure I could learn a lot from looking at prompts. As for the third question, there's no guarantee.

selridge 8mo ago
No the same question CANNOT be asked of source code because it can execute. You might as well ask for a record of the conversations between two engineers while code was being written. That's what the chat is. I have a pre-pre-alpha project which already has potentially hundreds of "prompts"--really turns in continuing conversations. Some of them with 1 kind of embedded agent, some with another. Some with agents on the web with no project access. Sometimes I would have conversations about plans that I drop. do I include those, if no code came out of them but my perspective changed or the agent's context changed so that later work was possible? I don't mean to be dismissive, but maybe you don't have the necessary perspective to understand what you're asking for.
- btrettel 8mo ago
  > maybe you don't have the necessary perspective to understand what you're asking for. I disagree. Thinking about this more, I can give an example from my time working as a patent examiner at the USPTO. We were required to include detailed search logs, which were primarily autogenerated using the USPTO's internal search tools. Basically, every query I made was listed. Often this was hundreds of queries for a particular application. You could also add manual entries. Looking at other examiners' search logs was absolutely useful to learn good queries, and I believe primary examiners checked the search logs to evaluate the quality of the search before posting office actions (primary examiners had to review the work of junior examiners like myself). With the right tools, this is useful and not burdensome, I think. Like prompts, this doesn't include the full story (the search results are obviously important too but excluded from the logs), but that doesn't stop the search logs from being useful. > You might as well ask for a record of the conversations between two engineers while code was being written. No, that's not typically logged, so it would be very burdensome. LLM prompts and responses, if not automatically logged, can easily be automatically logged.
  - selridge 8mo ago
    > LLM prompts and responses, if not automatically logged, can easily be automatically logged. What will you do with what you’ve logged? Where is “the prompt” when the chat is a chat? What prompt “made” the software? If you’re assuming that it is prompt > generation > release, that’s not a correct model at all. The model is *much* closer to conversations between engineers which you’ve indicated would be burdensome to log and noisy to review.
    - btrettel 8mo ago
      > What will you do with what you’ve logged? Could be a wide variety of things. I'd be interested in how rigorously a software was developed, or if I can learn any prompting tricks. > Where is “the prompt” when the chat is a chat? > The model is much closer to conversations between engineers which you’ve indicated would be burdensome to log and noisy to review. I disagree. Yes, prompts build on responses to past prompts, and prompts alone are not the full story. But exactly the same thing is true at the USPTO if you replace "prompts" with "search queries" and no one is claiming that their autogenerated search logs are burdensome. Also, the burden in actual conversations would come from the fact that such conversations are often not recorded in the first place. And now that I think about it, some organizations do record many meetings, so it might be easier than I'm thinking. > What prompt “made” the software? All of them.
selridge 8mo ago
I can't reply to the other comment, but here goes: This is one (1) conversation: https://chatgpt.com/share/69991d7e-87fc-8002-8c0e-2b38ed667384 https://chatgpt.com/share/69991d7e-87fc-8002-8c0e-2b38ed6673... It has 9 "prompts" On just the issue of path re-writing, that's probably one of a dozen conversations, NOT INCLUDING prompts fed into an LLM that existed to strip spaces and newlines caused by copying things out of a TUI. It's ok for things to be different than they used to be. It's ok for "prompts" to have been a meaningful unit of analysis 2 years ago but pointless today.

selridge 8mo ago

No the same question CANNOT be asked of source code because it can execute. You might as well ask for a record of the conversations between two engineers while code was being written. That's what the chat is. I have a pre-pre-alpha project which already has potentially hundreds of "prompts"--really turns in continuing conversations. Some of them with 1 kind of embedded agent, some with another. Some with agents on the web with no project access. Sometimes I would have conversations about plans that I drop. do I include those, if no code came out of them but my perspective changed or the agent's context changed so that later work was possible? I don't mean to be dismissive, but maybe you don't have the necessary perspective to understand what you're asking for.

btrettel 8mo ago
> maybe you don't have the necessary perspective to understand what you're asking for. I disagree. Thinking about this more, I can give an example from my time working as a patent examiner at the USPTO. We were required to include detailed search logs, which were primarily autogenerated using the USPTO's internal search tools. Basically, every query I made was listed. Often this was hundreds of queries for a particular application. You could also add manual entries. Looking at other examiners' search logs was absolutely useful to learn good queries, and I believe primary examiners checked the search logs to evaluate the quality of the search before posting office actions (primary examiners had to review the work of junior examiners like myself). With the right tools, this is useful and not burdensome, I think. Like prompts, this doesn't include the full story (the search results are obviously important too but excluded from the logs), but that doesn't stop the search logs from being useful. > You might as well ask for a record of the conversations between two engineers while code was being written. No, that's not typically logged, so it would be very burdensome. LLM prompts and responses, if not automatically logged, can easily be automatically logged.
- selridge 8mo ago
  > LLM prompts and responses, if not automatically logged, can easily be automatically logged. What will you do with what you’ve logged? Where is “the prompt” when the chat is a chat? What prompt “made” the software? If you’re assuming that it is prompt > generation > release, that’s not a correct model at all. The model is *much* closer to conversations between engineers which you’ve indicated would be burdensome to log and noisy to review.
  - btrettel 8mo ago
    > What will you do with what you’ve logged? Could be a wide variety of things. I'd be interested in how rigorously a software was developed, or if I can learn any prompting tricks. > Where is “the prompt” when the chat is a chat? > The model is much closer to conversations between engineers which you’ve indicated would be burdensome to log and noisy to review. I disagree. Yes, prompts build on responses to past prompts, and prompts alone are not the full story. But exactly the same thing is true at the USPTO if you replace "prompts" with "search queries" and no one is claiming that their autogenerated search logs are burdensome. Also, the burden in actual conversations would come from the fact that such conversations are often not recorded in the first place. And now that I think about it, some organizations do record many meetings, so it might be easier than I'm thinking. > What prompt “made” the software? All of them.

selridge 8mo ago

I can't reply to the other comment, but here goes: This is one (1) conversation: https://chatgpt.com/share/69991d7e-87fc-8002-8c0e-2b38ed667384 https://chatgpt.com/share/69991d7e-87fc-8002-8c0e-2b38ed6673... It has 9 "prompts" On just the issue of path re-writing, that's probably one of a dozen conversations, NOT INCLUDING prompts fed into an LLM that existed to strip spaces and newlines caused by copying things out of a TUI. It's ok for things to be different than they used to be. It's ok for "prompts" to have been a meaningful unit of analysis 2 years ago but pointless today.