Hey HN, I recently built a plan-validate-solve agent and struggled to write high-quality prompts so I decided to build a Grammarly for prompts. I used Go, HTMX, a-h/templ, Tailwind as well as the new OpenAI assistants - wanted to share a few learnings with you here:
- The overall “weight” of the solution. Basically, no external dependencies needed. I haven’t even touched web frameworks like Gin so far
- The packaging size when putting everything into an alpine-based image
- Super fast hot reload with cosmtrek/air in comparison to TS & React
- Basically no learning curve moving from React- to Templ-Tailwind templating
- Almost no JS is needed (except for interacting Web APIs like Permissions and Clipboard API)
For now, handling state on the server side is no big deal when using HTMX
So far I only faced some minor issues where, in some cases, Templ’s wrappers didn’t always transpile correctly or where hx-vals isn’t working due to Templ’s lack of support for single-quotes (used URL Query Params instead for now).
Re the OpenAI assistants - They worked quite well for my use case. But going forward I don’t see myself using them for very long since they are “just” isolated model instances that you can instruct and connect to your internal documents in a no-code fashion. I’ll probably drop them for a self-hosted Mistral 7B in the coming days, especially since fine-tuning becomes drastically easier these days. Excited to see how this scales!
An interesting direction I’ve been thinking about for the solution is to build something like this as a proxy that sits in front of LLMs and optimizes user prompts on the fly. Would appreciate any feedback on this approach.
- The overall “weight” of the solution. Basically, no external dependencies needed. I haven’t even touched web frameworks like Gin so far - The packaging size when putting everything into an alpine-based image - Super fast hot reload with cosmtrek/air in comparison to TS & React - Basically no learning curve moving from React- to Templ-Tailwind templating - Almost no JS is needed (except for interacting Web APIs like Permissions and Clipboard API)
For now, handling state on the server side is no big deal when using HTMX So far I only faced some minor issues where, in some cases, Templ’s wrappers didn’t always transpile correctly or where hx-vals isn’t working due to Templ’s lack of support for single-quotes (used URL Query Params instead for now).
Re the OpenAI assistants - They worked quite well for my use case. But going forward I don’t see myself using them for very long since they are “just” isolated model instances that you can instruct and connect to your internal documents in a no-code fashion. I’ll probably drop them for a self-hosted Mistral 7B in the coming days, especially since fine-tuning becomes drastically easier these days. Excited to see how this scales!
An interesting direction I’ve been thinking about for the solution is to build something like this as a proxy that sits in front of LLMs and optimizes user prompts on the fly. Would appreciate any feedback on this approach.