As humans, we like to share our original artifacts. When I take a great photograph, I might choose to run it through a compression algorithm to make it shareable, but I expect your software to then decompress it and help you see, more or less, the original.
There is an analogy here with using LLMs. While LLMs can be used as compression tools (“please summarize this document for me”), they are often used as decompression tools (e.g., “please make a CLI for Todoist”, or “please generate a report on the following dataset”).
When we use LLMs as decompression tools, it is good to share the resulting artifact, but it is also good to share the original prompt we used. Imagine a day when Web browsers have not only a view-source feature to see the source code but also a view-intent feature to inspect the text that generated the source!
In fact, to go a step further, those prompts are a kind of source — and that is true whether you are using the LLM to create software or any other artifact.1
Here are some benefits to sharing our prompts as a kind of source:
- We can save others the token costs of pointing an LLM at the resulting artifact and asking for the extracted intent. A lawyer friend joked that people are now sending him legal documents that Claude produced; he wishes they would just send him the prompt directly.
- People can learn things from how we prompt.
- “Forking” software at a higher level becomes possible. Maybe you have imagined an application that has 10x more features than I need, but is really thoughtfully designed; I could take your prompts, cut the parts I do not want, add something I do want, and ask my LLM to crank it out.
- Sharing prompts, instead of source, might provide a layer of obfuscation of the source code that is valuable in the current daily deluge of CVEs. Security vulnerabilities are found in the software supply chain, which is encoded in our source code, but not necessarily part of our prompts — I rarely specify desired dependencies in my prompts… if anything, I ask the LLM to minimize dependencies to maximize freedom in making changes.
And yes, I understand that we will not share all of our prompts with everyone — the prompts are, in some cases, the secret sauce. But when you can, share them. We might even create a new kind of open source with open prompts.
-
I feel we are lacking a good term for these prompts. There are some terms arising for these prompts in software, but prompts can be used to generate more than just software (we can make art, videos, music, fiction, etc. — I’m not making claims about the quality, but it will improve). I am looking for a term that communicates the weight of the prompt as a source, without communicating determinism or a software-only use case. ↩