github.com/oxremy/Blin...

When I ask it to write new CRUD logic, the composer is able to create six files in one go that are 90% good to go 🚀🤯

Additionally, for some tricky logic that touches multiple services/files, I asked Gemini to write an MD tutorial with examples (e.g., CRUD logic that involves front-end, backend, etc.). I also reference this MD file in the cursor rules (but sometimes add it to compose context manually).

Used LangChain.js with Aviation.bot. I'm also using Langchain Python to stream events (tool calls, etc.) to the front end. So in the future, I might be able to easily add more advanced customer service tools to the agent (e.g., look up an invoice, create a refund). 7/7

I liked the generated outputs so much that I also use them to draft customer support emails as part of the iteration effort and validation. Customers started replying more often with messages like "Thank you for the detailed and quick reply. 6/7

For reference, GPT-4 gave terrible answers and was much slower: ~30 sec vs 10 sec for Flash 2.0 generating a full answer. Google seems on the right track: faster, better and cheaper (for this use case) 5/7

It runs on a Cloudflare Worker + Langchain.JS + Gemini API. It pulls in the 150K token context using a markdown file that is stored in Cloudflare R2 (which is easily editable). The frond-end chat is just bootstrap + @levelsio style Jquery 🔥 4/7

Haven't seen any hallucinations. Iterated a lot on the system prompt by stress testing it and analyzing user queries for 2 weeks.
Need to improve the chat/code UI formatting a bit more. 3/7

BestTime.app API has a steep learning curve, and it's pretty good at suggesting the right API when the user explain's their use-case.

It can also give a cost estimate and helps the user pick a plan, and helps to users navigate to the right page (it returns hyperlinks in the chat). 2/7

Next step is to have an agent who lives inside my GitHub repository and automatically proposes a PR based on the Sentry issue when I wake up. I tried this last year with Sweep.Dev, but it wasn't smart enough. Any suggestions for new 2025 GH Repo AI tools?

Any feedback on the tone of voice, persuasion, or text itself? Maybe need to tweak the pauses and add a screen recording.

I still need to tweak the text and screen record the app itself.

@play_ht aviation.bot

Any feedback on the tone of voice, persuasion, or text itself? Maybe need to tweak the pauses and add a screen recording.

I still need to tweak the text and screen record the app itself.

Maybe it should have a Chrome Extension that sends page data to an MCP server, which then forwards the results to Cursor.

The current Cursor compose agent understands the codebase but lacks the crucial ability to track DB data as it flows into the code and gets rendered for the user. With more contextual rendering, agents could write and iterate on code much more effectively.

Any ideas or feedback? Should I encourage Cursor to do this automatically through .cursorrules?