I wrote up how I am using LLMs to help me code. It has been a lot of fun and quite mad. I really like it. Let me know your workflow! https://harper.blog/2025/02/16/my-llm-codegen-workflow-atm/ - ThreadSky

harper.lol • 125 days ago

I wrote up how I am using LLMs to help me code. It has been a lot of fun and quite mad. I really like it.

Let me know your workflow!

https://harper.blog/2025/02/16/my-llm-codegen-workflow-atm/

Comments

gosha.net•122 days ago

Thank you for writing this up, it’s super useful!

harper.lol•121 days ago

Glad it was helpful

raf.eco•123 days ago

Thank you for this post -- the practical ideas are great but what I really liked is that it captures how different it feels to really lean into the capabilities of these new tools.

deanputney.bsky.social•125 days ago

Very interesting. I’m going to have to try some of this!

The more things get established in my code base, the less useful LLMs appear to be. At some point autocomplete becomes the main use. I might be missing something…

harper.lol•125 days ago

Try it. I am having the opposite experience

danielraffel.bsky.social•121 days ago

Thanks for sharing! I (hope it's OK) I ended up making a Custom GPT from step 1 which I wrote about here: https://danielraffel.me/2025/02/21/building-a-custom-gpt-for-developer-ready-specs/

harper.lol•121 days ago

Amazing. I love it.

gulliegils.bsky.social•44 days ago

Definitely going to try this out.

tcjr.org•124 days ago

This is great, Harper. Thank you.

sstranger.bsky.social•123 days ago

Thanks for writing down your lessons learned 🙏

sammyrosen.bsky.social•125 days ago

wild!

just-john-j.bsky.social•28 days ago

Great writing and super useful. Thank you!!!

tomaslucas.bsky.social•44 days ago

Thank you for sharing your programming workflow. It aligns with others I follow, yet brings a fresh perspective on how to approach things using the new tools available.
1/3

tomaslucas.bsky.social•44 days ago

For example, I watch all the videos from this channel: https://www.youtube.com/@indydevdan/videos.
2/3

tomaslucas.bsky.social•44 days ago

I'd also like to recommend this interview with Boris Cherny and Cat Wu, where they discuss Claude Code: https://youtu.be/zDmW5hJPsvQ?si=_k1kSGdf4HW4AgQg
Lastly, congratulations on your presentation yesterday at the O'Reilly event — pure gold.
3/3

hopia.bsky.social•122 days ago

Martin Fowler like this, you better pay attention... Thanks Harper for the writeup!

adam-adair.bsky.social•44 days ago

I saw your talk on Oreilly today which led me to this blog post. Inspiring stuff! I don't have a workflow so I'm going to copy yours. Thank you!

harper.lol•44 days ago

Let me know how it goes

yocum137.bsky.social•125 days ago

📌

erkjlvk.bsky.social•121 days ago

This was an eye opener, thank you! One question: when working with large contexts, have you found ways to "cache" the context? Your test generation workflow seems to include the full repomix output each and every time round the loop? This could get expensive fast.

harper.lol•121 days ago

The repomix part is quick and cheap. And going to the llm can be token expensive. Aider helps by being more efficient. But context is key. By limiting context you are limiting capability.

I just lean into sharing the context as often as I can. It turns out be maybe every 10 turns.