Does anyone have great tips in using AI to help you craft prompts for claude-4 / cursor. 

I spend a lot of my time on the prompt and have had a hard time using AI to help refine them and improve. They tend to add scope and/or introduce confusion.

ive found with claude code I dont feel a need to be as formulaic with my prompting, just sharing as much context and being as clear about the goal as I can like I would with a person. adding claude.md files so you don't have to repeat context and having the agent share its execution plan + checklist + understanding of the tasks upfront has helped reduced the burden of making the initial prompt perfect

I know its not exactly what you asked for but as someone used to structuring my prompts strategically it's interesting to see how I'm using these tools change quickly

we just started using them so I'm still figuring out what's actually helpful but so far I've been using it dump documentation about specific parts of the code that I would otherwise need to repeatedly paste into the prompt.

for example on moshicam we have a photo editor with some complex gesture handling so I've documented all the available gestures, intended functionality and some edge cases the existing code is intended to handle. since you can put claude.md files in any directory, I feel like you can go pretty deep without overwhelming the context for unrelated tasks.

another example is with repetitive tasks like adding new notification types, there's boilerplate code that we need to add in several locations every time so having the steps outlined with examples helps it do it correctly without needing to explain it again. 

best part is that it can write the claude.md file for you so at the end of a task I'll ask it to document useful information and add any edge cases we encountered for future reference.

What do you usually put in the Claude.md file?

Engineer @farcaster, co-founder @bountycaster. Sharing what I learn about AI + development. Previously: Phantom, 0x. NYC 🗽

Dev ecosystem @farcaster, cofounder @bountycaster. Prev: cofounder of Scalar Capital, product at Coinbase, coproducer @ethereumfilm

Another effective approach:
- Run a deep dive and create a document outlining best practices for prompting (ideally tailored to your specific use case)
- Then feed that document into another LLM (Opus or o3) and ask it to refine your prompt using those best practices
Tip: Ask it to focus on the 80/20 rule, so it optimizes the key parts without completely rewriting your original prompt

@jacob have you tried Anthropic's prompt refiner? I remember it being pretty good

"hell no, I don't vibecode! I take pride in my craft and write things myself" said a devrel at cannes. I'm at the opposite end of the spectrum using LLMs all day and almost don't look at code anymore. here's what I do

hellno vibe coding snapshot july 2025
· use subtasks in parallel: overall speedup and LLMs are good at deciding what should be done in sequence vs at the same time in different processes

· get context and write lean PRDs: force the LLM to think longer improves everything. I was hesitant to do this after seeing @jc4p and @stephancill casting about it. turns out they are so right! taking time before writing code increases output quality A LOT and saves time to not get stuck in bug fix-vibe debugging loops later. 

· increase input speed
A) use snippets to replace keywords into expanded prompts. tiny example: /par → continue, execute subtasks in parallel if possible 
B) speech to text is fun and much faster. might require setting up some dictionary to recognize some words, but overall fast and easy

posted my prompt flow and github link to more prompts here. I'm usually creating a very lean PRD to avoid confusion and enforce scope

posted my prompt flow and github link to more prompts here. I'm usually creating a very lean PRD to avoid confusion and enforce scope
https://farcaster.xyz/hellno.eth/0x124d02cb

What is wrong with the results you’re getting currently?

https://docs.anthropic.com/en/docs/build-with-claude/prompt-engineering/overview

More recently all I do is wrap large context blobs in xml tags

What are you hoping the prompt enables that it failed at before

When it struggle I try to text using two languages, for example, mix portuguese and english on the prompt. It helps a bit, at least for me

Did you tried others AI programming tools?

Keep it tight: “Refine for clarity, no added scope. Give 3 sharper versions.”