narphorium.bsky.social
Building AI powered tools to augment human creativity and problem solving in San Francisco. Previously @GitHub Copilot, @Google, 🇨🇦
169 posts
1,279 followers
556 following
Regular Contributor
Active Commenter
comment in response to
post
In Cursor and in Claude chat
comment in response to
post
I've found that Gemini 2.5 pro and Claude 4 both do a much better job of pushing back on design decisions that they disagree with
comment in response to
post
I think, in the past, I would have felt that I needed to package them up as libraries but now they're just a bunch of little scripts and Jupyter notebooks.
comment in response to
post
There's something really interesting about that “micro pipeline” pattern. I find myself writing a lot more little fragments of useful code now that I know that the cost of integrating them back into a larger project is almost nothing.
comment in response to
post
I have a custom brainstorming prompt that I use to help me expand on ideas and Claude 4 is now challenging the underlying assumptions in ways that make me have to sit back and really think for a moment
comment in response to
post
Haha, yeah. I've noticed AI coding tools let me keep coding way past the point where I should have stopped
comment in response to
post
It’s their “copilot” in more ways than one
comment in response to
post
Saw one of them in SF this week. Very weird design
comment in response to
post
I would love to see something like this with natural language constraints like “X should be a above Y”
comment in response to
post
Make sure to say “context is king” 😄
But seriously, please post it. This is one area where I think a lot of people writing about the same thing is valuable
comment in response to
post
Yeah its hard to get started but you can ask an AI agent to read through one of your hand coded projects to build a style guide; then iterate on that
comment in response to
post
docs.cursor.com/context/rules
comment in response to
post
docs.cursor.com/context/rules
comment in response to
post
docs.github.com/en/copilot/c...
comment in response to
post
Did you try any custom rules to teach it your coding style?
comment in response to
post
comment in response to
post
I've been exploring some similar interactions with reasoning models bsky.app/profile/narp...
comment in response to
post
I'd love to see more people building tools like this. So far I haven't found any tools which tell me how my custom rules are being used
comment in response to
post
Thanks! It goes through the chat history one-by-one and asks the model which rules are relevant to the conversation and which rules are redundant given the existing context. Then all of the results get combined into a final set of suggested edits based on how relevant or redundant each rule is.
comment in response to
post
You should check out "Latent Patterns in Activities: A Field Study of How Developers Manage Context" which looks at how human developers decompose tasks.
rahulpandita.github.io/files/riniIC...
comment in response to
post
Congratulations!
comment in response to
post
It's still very much in beta, but I'm really excited about the flywheel effect this creates: each agent interaction informs better rules, and a growing library of personalized rules creates a faster, more effective coding assistant.
6/6
comment in response to
post
I based the algorithm on techniques from model self-evaluation research so that the LLM reads through my chat history and its suggestions are grounded in how the custom rules are being used.
arxiv.org/abs/2306.05685
5/6
comment in response to
post
Unfortunately, most tools treat custom rules as a black box. You input instructions and hope for the best, with no visibility into what's actually happening.
4/6
comment in response to
post
But how am I supposed to know if my instructions are helping or hurting? What I needed was something like stacktraces for natural language - a way to see exactly how my instructions are being interpreted, where they're failing, and why.
3/6
comment in response to
post
Agents can dramatically accelerate software development but acceleration in the wrong direction can be worse than no acceleration at all. I believe the key to keeping agents focused on the right tasks will be higher quality, more personalized instructions.
2/6