Quentin Fennessy

this is insane. at a cost of $2.60 #claudecode improved accuracy of the prompt from 26% to 94%, hands off, by running tests, reviewing evals, and updating the prompt. #claudecode even added an option to limit test run size, which is a cost optimization. I have not even eaten breakfast yet.

this morning I found more #claudecode magic. I created eval jsons, with sentences and expected extracted data, with #claude. I had #claudecode then create an extraction prompt and an evaluation script to score the results. Right now #claudecode is iterating on the prompt, by itself. OMG.

#anthropic has kindly resolved the pytest bug with #claudecode. thank you very much!

#claudecode is tired of my demands. This morning I got this:
"Looking at the errors, your tests are mocking the client's response but the repository methods aren't actually being called...
You need to modify your tests...
Try removing the client.get mocking ..."
Wait, what? no, plz fix

okay, in #claudecode you type /bug and report the, well, the bug!

the fun continues. I should figure out how to report this to #anthropic although I suspect they already know.

In many cases when I ask #claude of various generations "advise me" I need to research the enhancements that are suggested. Almost always very good ideas. Many times I would not have considered the improvement.

I have #claudecode create a plan, execute the plan in phases, commit code in phases with excellent commit messages, and then update the plan. This means I can see one md file in my repo that shows progress on work and deliverables. This amazes me.

at this point I have 172 files, 44k loc, and 750 pytests, giving me 49% test coverage. I'm trying to balance test coverage with speed of development.

perhaps I should have 3 local copies - one for integration, and two for development. I'm mixing and matching #claudecode and #windsurf because while I love the #claudecode interface I also love the tools and view provided by #windsurf.

I have a single repo. should I install two copies locally and have the two agents/bots/imaginary friends work on different branches? My first experiment has #claudecode and #windsurf working on the same repo in the same branch, on distinct parts of code. who else has tried thiss?

I am experimenting with a team of coding agents. I have #claudecode enhancing my api for performance, and at the same time I have #windsurf designing test coverage for my firebase services. Am I an indy hacker manager of bots?

#claudecode is confused w/ pytest-since yesterday: "However, we're still encountering issues with pytest itself, which appears to be a Python environment issue. The error message TypeError: _warning_redirect() missing 1 required positional argument:.." Not fatal as I can run pytest and validate.

#claudecode tip: when working on a significant enhancement, have Claude Code create the design. Review and modify it. Have Claude create an implementation plan. Claude should managed these docs in md files in your repo. And you can ask Claude to update the plan with progress. Works!

I've been very productive with Claude Code since release. Here's a tip that may be helpful: allow Claude Code to generate commit messages. they are very good. This gives you the history you need! #claude #claudecode

Never heard of it

is there guidance on prompting @windsurfai.bsky.social to preserve the underlying software structure while making changes? When I add a command to the CLI, and the related database operation, and the related test, I want it to follow my standard. how do I do that?

I'm using #Claude and #Windsurf in a new way (for me). First I ask Claude to break down a code change for me into steps. then for each step I ask Claude to create a prompt for that change. and then I paste that prompt into windsurf. so far it is working very well.

Open-R1: a fully open reproduction of DeepSeek-R1 We’re on a journey to advance and democratize artificial intelligence through open source and open science.

The most impactful open-source project of today (dixit Vercel VP of AI)
=> huggingface.co/blog/open-r1

It is really interesting that you could get that quality of a response from a 1B model. however I don’t think we’ve define what think means and how do we ask the question if we don’t know what think means?

I am learning the Temporal Python SDK This may be the simplest example of creating and managing workflows. This is an easy intro, and @temporal.io provides much more sophisticated code to continue learning form.

👉 www.qfennessy.net/posts/tempor...
👉 github.com/qfennessy/ba...

#Python #Temporal

Looks good. Will you share your code?

Arthurian is an experiment that creates shapes and patterns with a small range of styles.

inspired by:
@shurastei.bsky.social
@jeffpalmer.bsky.social
@pb67art.bsky.social
@whitekoalaart.bsky.social
Code freely available. github.com/qfennessy/ar...

#genartclub #creativecoding

This sounds like excellent advice. I will try this soon. Thank you.

Very very nice. Consider this reply as an homage.