You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
👉🏽 Consider the job writing guide in apollo's job_chat. Can we use it? Improve it? 👉🏽
I found the experience of using claude cli easier and best suited for local development. I don't think we should make openfn cli behave like claude code cli. It's not time well spent. i think openfn init should give you a project with a .claude file incase you want to use claude for job writing.
So what i have done is ask AI to create a claude.md based on apollos model and it created about 432 lines of instructions. And from there i could prompt claude to write openfn jobs easily. See my video here
👉🏽 How much detail do we need? Is one file enough?
I would say we start with 1 file and expand as we need. I think from using claude code we will know what to fine tune
But i also like the deaiify.md file from Lightning, See file here.I think it reduce that AI jargon. So maybe 2 files
👉🏽 Should we save claude.md in github and have the CLI clone it on openfn init?
I was thinking in project templates repo, But i think in kit it's the best home if we want to keep updating the file
👉🏽 Should we flip the script and use the claude.md to power our job chat?
Too soon to tell i would say, I think the thing i liked about using claude code is speed and the cli ux is good 👍🏽 ❤️ . But i am not convinces we should flip the switch just yet. Especially for longevity what is the actual cost of developing using claude code vs openfn ai ? Maybe we should compare API token usage between two approach 🤷🏽
👉🏽 Can we give it any access to external resources, like our doc site?
Yes, so in my claude.md I have a resources section which points to different documentation links and i asked claude to point me to the documentation for each function that was used and returned links to each and every function, even docs for lazy state was returned 🤯
Rather than creating a big claude.md file, let's just make a really good md job writing guide. We should then feed this to humans, claude, the CLI, and any other agents who l might be interested. Including the AI assistant maybe?
Activity
mtuchi commentedon Oct 17, 2025
Hiya @josephjclark, This is my feedback:
I found the experience of using claude cli easier and best suited for local development. I don't think we should make openfn cli behave like claude code cli. It's not time well spent. i think openfn init should give you a project with a .claude file incase you want to use claude for job writing.
So what i have done is ask AI to create a claude.md based on apollos model and it created about 432 lines of instructions. And from there i could prompt claude to write openfn jobs easily. See my video here
I would say we start with 1 file and expand as we need. I think from using claude code we will know what to fine tune
But i also like the deaiify.md file from Lightning, See file here.I think it reduce that AI jargon. So maybe 2 files
I was thinking in project templates repo, But i think in kit it's the best home if we want to keep updating the file
Too soon to tell i would say, I think the thing i liked about using claude code is speed and the cli ux is good 👍🏽 ❤️ . But i am not convinces we should flip the switch just yet. Especially for longevity what is the actual cost of developing using claude code vs openfn ai ? Maybe we should compare API token usage between two approach 🤷🏽
Yes, so in my claude.md I have a resources section which points to different documentation links and i asked claude to point me to the documentation for each function that was used and returned links to each and every function, even docs for lazy state was returned 🤯
TL;DR Here is the video walkthrough of my research
josephjclark commentedon Oct 20, 2025
Discussion points:
mtuchi commentedon Oct 20, 2025
Hiya @josephjclark i have created a general job writing guide for ai services, See OpenFn/project#18