Matthew pfp

Matthew

@matthew

3180 Following
142415 Followers
Matthew pfp
0 reply
0 recast
3 reactions
Matthew pfp
6 replies
1 recast
39 reactions
Matthew pfp
3 replies
1 recast
9 reactions
Matthew pfp
0 reply
0 recast
14 reactions
Matthew pfp
initial thoughts after hacking with claude code (after using only cursor for 6 months) 1. I weakly prefer the output from Claude code, which is interesting because I use sonnet 4 and opus exclusively in cursor. It seems to be stronger in task management (or at least in its communication of it), which I don't notice as much from cursor agents. 2. strongly prefer cursor's visual UI vs terminal, even CC uses the terminal about as well as it could 3. not sure if cursor has an "auto accept all edits" mode but if so I haven't found it. I like that Claude code does this, but I do wish it was easier to evaluate the work it did. Feels like OpenAI was smart to acquire windsurf, because I find myself wishing for Claude code + Cursor. Which maybe is just Cursor in a few months? Anyway... on to try @mcbain's amp code
3 replies
1 recast
10 reactions
Matthew pfp
3 replies
1 recast
13 reactions
Matthew pfp
6 replies
2 recasts
35 reactions
Matthew pfp
3 replies
0 recast
3 reactions
Matthew pfp
0 reply
0 recast
14 reactions
Matthew pfp
3 replies
0 recast
18 reactions
Matthew pfp
1 reply
0 recast
12 reactions
Matthew pfp
One really useful way that AI has changed my coding workflow is that now I rarely start constructing a component, file, system, UI, etc from scratch. I'll usually give Claude Opus a ton of context and have it write a good v1, then go in and rewrite it myself or with AI so that it works properly, considers all edge cases, looks right, etc. A recent example of this is a component we're in the process of implementing now called <TokenBalance />. It's an easy drop-in component that takes a contract address, account address, and chain id, and produces a consistently formatted and stylized balance for that token—USD value + token amount + token ticker. I spent a week on-and-off writing a full spec, getting feedback from other engineers, and tweaking it to exactly what I wanted. Once it was bulletproof, I had Claude write the v1. there were a few small things it got wrong, but it was mostly correct, and soooooo much easier than building it myself from scratch. I went through and tweaked various things, and each time had the AI tweak all my examples and documentation to match my changes. Again—so helpful to eliminate much of the effort of staying consistent. All in all the building part of this took about a day, easily 2-3x faster than I would've been able to do on my own. And most of my involvement was in reviewing the work of the AI, rewriting things it got wrong, judging the difference in variations and tweaking to match my taste, etc.
6 replies
3 recasts
42 reactions
Matthew pfp
1 reply
1 recast
7 reactions
Matthew pfp
0 reply
0 recast
21 reactions
Matthew pfp
3 replies
1 recast
6 reactions
Matthew pfp
1 reply
0 recast
12 reactions
Matthew pfp
5 replies
0 recast
36 reactions
Matthew pfp
10 replies
0 recast
7 reactions
Matthew pfp
3 replies
0 recast
14 reactions
Matthew pfp
1 reply
0 recast
4 reactions

Mini Apps

Eventcaster - Farcaster Mini Apps

A Farcaster Mini App by matthew

A mini app by matthew.

Poker - Farcaster Mini Apps

A Farcaster Mini App by matthew

Play poker with friends