Pedro Figueiredo
@P_FFigueiredo
Followers
215
Following
1K
Media
63
Statuses
583
29, Member of Technical Staff @ Databricks, TypeScript, State Machines & AI.
Joined December 2013
We've been experimenting with Claude Skills. Yes, they're just markdown - but the interesting thing about Skills isn’t their anatomy, it’s the way Claude manages them:
1
1
24
Cursor's new layout just seems like "natural evolution" for me. It's putting natural language as the first class citizen, instead of the actual code, love it 🔥
1
1
12
Right now, it only supports Claude Code through their SDK, but more is coming pretty soon! Cool fact, this is how we are testing Claude Code Skills for Neon 👀
0
0
0
“Getting started with AI” is becoming super popular! That’s usually a prompt or some tools you can feed Agents with and get started coding in a nutshell. To make sure those prompts/tools work with those agents, I’m building: https://t.co/zWPbXkOCz8 🚀
github.com
TypeScript library for evaluating prompts against coding agents (Claude Code, Cursor, etc.) with multi-iteration testing and scoring - pffigueiredo/code-agent-eval
1
0
0
Working with AI and cutting edge tehcnology is not easy, I can tell you that much. You really need to keep reminding whatever agent you are using to look at specific documentation, since keeping it all at context seems to be even more inneficient.
0
0
1
100% this. Custom "plan builders commands" are still needed today, the built-in plan modes aren't enough to build complex and complete features in my experience.
more context engineering == you can ship bigger features in spicier codebases
0
0
0
The `haiku-4-5` models feels so damn good for dumb work or implementing detailed plans 😱
0
0
0
This is what I'm thinking. If the MCP server you are using doesn't require authentication, it should be a "claude skill" instead. The reason is simple, much less noise in the context window for about the same value.
1
0
1
With the Haiku 4.5 release, there is now an opportunity to change a lot of sub-agents default model, at least for the ones that do code discoverability and where the quality generation is a bit less important.
1
0
0
If you are an engineer, I think one of the worse things you can do when working with AI, is delegating the entire thought process to it. Use AI to generate the code, not to create the solutions for your problems. 👈
0
0
0
Looking to learn a bit more about Claude Code and how to use it to do actual engineering work? Here is a good start: https://t.co/QbGetrDGuH More to come very soon! 😎
neon.com
Claude Code is part of our daily workflow at Neon. Here’s a practical cheatsheet of the features, commands, and tweaks we rely on most.
1
0
1
I've introduce 2 new small things to my CC workflow: 1. An /update-documentation command 2. A new status into my bar to let me know when my docs are unsynced with the code You can keep track of this in: https://t.co/jjeU4hm4GK
0
0
0
Something that I thought was cool from Sonnet 4.5, is that it manages it's own context window, by compacting the conversation when it is close to the token limit. However, Cognition already had to re-write part of their product because of that! 👇 https://t.co/eMMwm1URnm
cognition.ai
We rebuilt Devin for Claude Sonnet 4.5. The new version is 2x faster, 12% better on our Junior Developer Evals, and it's available now in Agent Preview. For users who prefer the old Devin, that...
0
1
3
Just got back from two weeks vacation, and although I'm a bit rusty, AI seems to be have gone better, so, I will take the trade-off for now 😀
0
0
2
I love seeing prompting frameworks like `spec-kit` getting some traction. However, most seem over-engineered and will waste too many tokens and time for simple tasks. It can easily take up to 40 minutes to run trhough this flow: -> specify -> plan -> tasks -> implement
2
0
3
This week I'm switching from Claude Code to Cursor. I'm doing this for these reasons: - understand how locked-in I'm to anthropic tooling - checking new Cursor features - writing a comparing blog post
1
0
4