primaprashant

Been working on a weekly newsletter [1] to stay fully informed about agentic coding with one email, once a week. I also keep the focus narrow, only on what engineers and tech leaders would find useful for shipping code and leading teams, which means I filter out all generic AI news, or what CEO said what, or any marketing fluff.

[1]: https://www.agenticcodingweekly.com/

They used this exact treatment in an episode of The Good Doctor, S01E06. Original air date: October 30, 2017

https://the-good-doctor.fandom.com/wiki/Not_Fake

You should check out Devstral 2 Small [1]. It's 24B and scores 68.0% on SWE-bench Verified.

[1]: https://mistral.ai/news/devstral-2-vibe-cli

yeah, big fan of uv and ruff and recently ty. It's also from astral and it fixed type checking. You should check it out

They also have an API which you can use to get the icon SVG.

I love making (architecture) diagrams in D2 [1], and love using the vast library of icons from Iconify in my diagrams where it makes sense. A sample diagram with SVG from Iconfiy would look like this:

  docker: Docker {
    icon: https://api.iconify.design/logos/docker-icon.svg
  }

  kubernetes: Kubernetes {
    icon: https://api.iconify.design/logos/kubernetes.svg
  }

  docker -> kubernetes: deploy

[1]: https://d2lang.com/

This is really nice. I keep track of most important habits to me like how often I go to gym, how much protein I eat everyday, and how many days I read (books), on something physical (pen and paper). Mostly on monthly calendars. This would make tracking each of them separately on a single piece of paper across the entire year pretty neat.

Started making side projects as a developer this year and hope to start working on my own products full-time from next year. Two books I found useful for positioning the product:

- Obviously Awesome by April Dunford (https://www.goodreads.com/book/show/45166937-obviously-aweso...)

- Building a StoryBrand by Donald Miller (https://www.goodreads.com/book/show/210137279-building-a-sto...)

Anyone specifically looking for ML engineering blogs should find this useful: https://github.com/primaprashant/ml-engineering-blogs

Good stuff. You will enjoy my short essay, I want to give a lot of fucks! [1], which argues against the typical conclusion reached by people working at big corp long enough: "Stop caring. Stop giving a fuck. Focus on things outside of work".

The core insight it, if you start to feel the need to stop caring, instead of changing your character and values, treat it as a strong signal to change your environment.

[1]: https://anandprashant.com/posts/i-want-to-give-a-lot-of-fuck...

Pricing is $0.5 / $3 per million input / output tokens. 2.5 Flash was $0.3 / $2.5. That's 66% increase in input tokens and 20% increase in output token pricing.

For comparison, from 2.5 Pro ($1.25 / $10) to 3 Pro ($2 / $12), there was 60% increase in input tokens and 20% increase in output tokens pricing.

Started a newsletter [1] focused on agentic coding updates, nothing else. Other newsletters/blogs cover a lot of generic AI news, industry gossip, and marketing fluff. Having a focused feed is something I wanted for myself and finally I have enough time that I can write this newsletter regularly.

[1]: https://www.agenticcodingweekly.com/

I built my own open-source tool to do exactly this so that I can run something like `claude $(hns)` in my terminal and then I can start speaking, and after I'm done, claude receives the transcript and start working. See this workflow here: https://hns-cli.dev/docs/drive-coding-agents/

Created a summary of comments from this thread about 15 hours after it had been posted and had 814 comments with gemini-3-pro and gpt-5.1 using this script [1]:

- gemini-3-pro summary: https://gist.github.com/primaprashant/948c5b0f89f1d5bc919f90...

- gpt-5.1 summary: https://gist.github.com/primaprashant/3786f3833043d8dcccae4b...

Summary from GPT 5.1 is significantly longer and more verbose compared to Gemini 3 Pro (13,129 output tokens vs 3,776). Gemini 3 summary seems more readable, however, GPT 5.1 one has interesting insights missed by Gemini.

Last time I did this comparison at the time of GPT 5 release [2], the summary from Gemini 2.5 Pro was way better and readable than the GPT 5 one. This time the readability of Gemini 3 summary still seems great while GPT 5.1 feels a bit more improved but not there quite yet.

[1]: https://gist.github.com/primaprashant/f181ed685ae563fd06c49d...

[2]: https://news.ycombinator.com/item?id=44835029

Building a docs website [1] for my speech-to-text CLI tool, hns. I use it 5-10 times daily to transcribe my voice, and a few developer friends I've shared it with have also adopted it for daily use. They like that it runs in the terminal and keeps all data local. So, felt like I should write down guides for new users to get started quickly and to highlight key use cases.

Building this documentation website using Docusaurus. This is my first time using Docusaurus, and it feels like a very nice tool for quickly developing a documentation website.

[1]: https://hns-cli.dev/

built something similar for terminal lovers. It's a CLI tool built in Python called hns [1] and uses faster-whisper for completely local speech-to-text. It automatically copies the transcription to the clipboard as well as writes to stdout so you seamlessly paste the transcription in any other application or pipe/redirect it to other programs/files.

[1]: https://github.com/primaprashant/hns

btw, if you want local dictation, speak and get a transcript, not transcribe files, I built a Python tool called hns [1]. It's open source, uses faster-whisper, and you can run it with `uvx hns` or just `hns` after `uv tool install hns`.

[1]: https://github.com/primaprashant/hns

Only have access to GPT-5 through API for now. The amount of tokens (>130k) used is higher than the limit of ChatGPT (128k) so it wouldn't really work well.

yes, agreed. Context length might be playing a factor as total number of prompt tokens is >120k. Performance of LLMs generally degrade at higher context length.

created a summary of comments from this thread about 15 hours after it had been posted and had 1983 comments, using gpt-5-high and gemini-2.5-pro using a prompt similar to simonw [1]. Used a Python script [2] that I wrote to generate the summary.

- gpt-5-high summary: https://gist.github.com/primaprashant/1775eb97537362b049d643...

- gemini-2.5-pro summary: https://gist.github.com/primaprashant/4d22df9735a1541263c671...

[1]: https://news.ycombinator.com/item?id=43477622

[2]: https://gist.github.com/primaprashant/f181ed685ae563fd06c49d...

looks like 4 new features for API

- reasoning_effort parameter supports minimal value now in addition to existing low, medium, and high

- new verbosity parameter with possible values of low, medium (default), and high

- unlike hidden thinking tokens, user-visible preamble messages for tool calls are available

- tool calls possible with plaintext instead of JSON

GPT-5 was supposed to make choosing models and reasoning efforts simpler. I think they made it more complex.

> GPT‑5’s reasoning_effort parameter can now take a minimal value to get answers back faster, without extensive reasoning first.

> While GPT‑5 in ChatGPT is a system of reasoning, non-reasoning, and router models, GPT‑5 in the API platform is the reasoning model that powers maximum performance in ChatGPT. Notably, GPT‑5 with minimal reasoning is a different model than the non-reasoning model in ChatGPT, and is better tuned for developers. The non-reasoning model used in ChatGPT is available as gpt-5-chat-latest.

Building a local CLI tool, hns [1], for speech-to-text in terminal in 5 keystrokes.

It records my voice, transcribes it locally using faster-whisper, and copies the transcription to my clipboard. Check the demo linked in the GitHub repo to see how it works.

I use it especially with Claude Code to provide detailed context for the outcome I want to achieve. I ramble for 5 minutes, and then paste the transcription to Claude Code, instead of having to type all my thoughts all the time.

The workflow is like this:

  $ hns # start recording
  <talk>
  <Enter> # clipboard now holds the text

[1] https://github.com/primaprashant/hns

A minimal website (https://lgtms.app/) to get a random ASCII art for the acronym LGTM cause I love using them in the comment field while approving PRs. Supports CLI as well.

Tell me you know nothing about machine learning without telling me you know nothing about machine learning

Some cities already have this kind of community which is pretty active so people don't see the point of signing up to this. For example, I'm in Tokyo we have an almost monthly HN meetup, HN Tokyo [1].

[1]: https://hntokyo.doorkeeper.jp/

I use the self-hosted version of IT-Tools [1] on my own subdomain. I find it helpful that I can access these tools no matter which machine I'm using (as long as I have an internet connection).

[1] https://github.com/CorentinTh/it-tools/

> An analysis of the correlation between per capita coffee consumption and life expectancy across countries indicates a moderately positive correlation of around 0.577, suggesting there exists a slightly strong correlation between the two metrics.

> Due to the moderately strong correlation, performing RA on those metrics suggests that each additional kilogram of coffee consumed per person annually could potentially raise life expectancy by 1.22%.

I would imagine this to be a "correlation is not causation" situation. Countries with high coffee consumption per capita are generally developed countries.

I have found Claude 3.5 Sonnet really good for coding tasks along with the artifacts feature and seems like it's still the king on the coding benchmarks

The resources for link to model card[1], research paper, and Prompt Guard Tutorial[2] on the page doesn't exist yet

[1]: https://github.com/meta-llama/llama-models/blob/main/models/...

[2]: https://github.com/meta-llama/llama-recipes/blob/main/recipe...

Was the candidate informed about the expectations before or after being hired? Has anyone communicated to them that they are not meeting those expectations?