What people actually do
Each note is a public post: someone describing a task they did with a listed tool. The line is a short account of that post, with a link back to it. This is not a ranking, and a prompt is shown only when the source includes one.
49 reports · newest post 9 Oct 2026 · Job guides
On the Whistle speech-to-text thread, a commenter said Opus would not help set it up and that they used Grok to drive the setup.
An engineer said they used Claude over several months to implement a lossless compression idea, including medical imaging, and that it was still about 5% behind WebP on a Kodak test.
The commenter asked Grok to compile some Emacs Lisp into Vim script and reported that it was possible but strange.
The author used ChatGPT to investigate a YouTube bug that skipped backward in a video during soft reloads.
Wanting a cowboy-hat drawing for a birthday card, the poster used ChatGPT image generation to break the drawing into steps, then made a daily drawing-lesson site.
Panos Ipeirotis pasted a Norse Atlantic overnight-delay email into ChatGPT Pro, followed its EU261 and Norwegian-conciliation steps, and wrote that Norse later wired $4,760.36.
The author built a Rust and Bevy space-economy sim, with a few hundred ships running their own planners, and said they built it with Claude.
Sankalp used Codex in GPU Mode's auto-research contest to speed a batched QR kernel to 232x the baseline and finished 12th of 183.
Making an animation with about 40 lines, the poster said listeners found the ElevenLabs character voices flat and inconsistent.
On the Gravity solar-system simulator thread, a commenter said they used ElevenLabs for both the narrator and the music.
A commenter on an Ask HN about flagged Show HN posts said they used Grammarly because English is not their first language, and that they wrote the post themselves.
A commenter said they use Grammarly heavily on speech-recognition drafts because it catches misheard words and places where they thought one word and said another.
On a thread about ChatGPT for Excel, a Microsoft 365 Premium subscriber said they have had a good time with Copilot in Excel and tend to pick Opus 4.6.
David Connolly spent a few weeks with Claude Code on FFmpeg's long-open subtitle-conversion limit, and wrote that Claude produced most of the encoder and found an integer overflow in the decoder.
On a Show HN for an Arabic spaced-repetition app, the poster said they used Artlist and ElevenLabs for the audio and hope to replace it with native-speaker recordings.
While writing a first neural network, the poster used Gemini instead of Wikipedia to fill gaps, including learning dual numbers rather than integrating a computer-algebra library.
A refinery logistics manager who is not a software developer relied on Claude, Copilot, and Gemini to write a 9,000-line browser game explaining the plant to his kids.
A competitive-intelligence worker built about 15 Make.com scenarios to pull sources, filter them, summarize them with GPT, and email the results, then wrote that the scenarios later broke silently.
Eric Wagoner built an n8n workflow that collects his GitHub and Jira activity and emails a standup update, and wrote that the daily morale lift was the unexpected part.
Simon Willison ported the JustHTML HTML5 parser from Python to dependency-free JavaScript with Codex CLI and GPT-5.2 in about 4.5 hours.
Prompt they used
We are going to create a JavaScript port of ~/dev/justhtml - an HTML parsing library that passes the full ~/dev/html5lib-tests test suite. It is going to have a similar API to the Python library but in JavaScript. It will have no dependencies other than raw JavaScript, hence it will work great in the browser and node.js and other environments. Start by reading ~/dev/justhtml and designing the user-facing API for the new library - create a spec.md containing your plan.
Jeremy used a new Claude Max subscription and Claude Code to write a piano web app after a year of occasional OpenAI help on specific problems.
The poster said they rolled Microsoft Copilot out to 4,000 employees in the previous quarter.
In two weeks the poster used Gemini 3 to turn 42 books into 42 webpages, each meant to show the feel of that book.
Taylor illustrated and laid out the children's book Tech Is Magic in Canva Pro, using Magic Media for the pictures and Canva for the pages, without hiring an illustrator.
Ella Flitcroft made a yellow website header in Canva for a class graphics reflection, the first header she had built there.
After almost two years, William described using Perplexity for questions such as why dogs on an adoption site are named, whether $250 was fair for a used LG UltraFine 5K, how to use Stytch server-side with Next.js, and what a Visa Flexible Credential is.
Sanity staff engineer Vincent Quigley described running several Claude Code agents like a small team with daily amnesia, and said AI was writing about 80% of initial implementations.
Ben Lampere self-hosted n8n on a DigitalOcean Docker droplet and, from a Discord bot, ran subdomain enumeration, directory enumeration, and Gowitness screenshots back into Discord.
Responding to feedback on a post about vibe coding careers, the author said they mostly used Grammarly and that it caught small mistakes such as “remenber”.
Grok 4 helped the commenter trace inconsistent lldb-via-Python behavior between Docker and a local Linux box to a difference in how the address sanitizer worked in the two environments.
The author refactored the old Xavibot project almost entirely in Cursor's agent mode, including a chatbot backend move from Azure to Google Cloud, using Gemini 2.5 Pro, Claude 4 Sonnet, and often Auto.
Stuck in Bangalore traffic, the poster used Replit on the commute to build Top Agents, a small directory for searching and reviewing public AI agents.
Michal Pleban used DeepSeek V2 Chat for the ideas, prompts, descriptions, and keywords behind 464,708 tattoo designs, with Flux generating the images.
Dontizi posted RLAMA, an open-source document tool whose model list includes local models served by Ollama.
vikvang wrote that they won Robinhood's $1M trivia contest after building with Perplexity and OpenCV.
Denis Bratchikov used Cursor Composer, with GPT-4o and Claude 3.7 Sonnet, to strip deprecated Playwright arguments from 64 spec files.
Prompt they used
In `application\e2e`, find all `.spec.ts` files and do the following: 1. Remove object with `shouldClip: true` 2. Remove the empty string before the removed object (if any) 3. Remove `'new-ui-design': true,` from the corresponding object 4. If the corresponding object (e.g., `featureFlags`) becomes empty, remove it 5. If the parent object of the corresponding object becomes empty, remove it
Data journalist Shijith used Gemini 1.5 Pro to read Delhi election manifesto PDFs and draft a comparison across 10 issues, including pollution and education.
The poster used DeepSeek to turn Theo's Stripe setup notes into a CLI and wrote that they did not write any of the code themselves.
Evan Widloski forwards real-estate SMS spam over MQTT to a workstation running Ollama with llama3.3, and saves reply personas on that local model.
Evan Widloski used Cursor to change an Android SMS gateway so a received text becomes an MQTT message on a connection the app already had.
Prompt they used
[SMSListener.java (41-46)] How do I turn this POST request into an MQTT message? I'd like to use the MQTT connection already present in the file `BackgroundService.java` if possible.
The Show HN author posted an ASCII camera web app and said they made it with Replit Agent.
Yiannis Demetriou posted a tutorial on automating lead generation with Apify and Make.com, for an AI agency's own leads or to sell as a service.
The poster shared a video of a LangChain agent they built to chat with Hacker News.
Senna Parsa used GitHub Copilot Chat to prototype a React photo gallery of tulip fields and flower shows around Amsterdam for a conference demo.
IBM and NASA announced that IBM's watsonx.ai geospatial foundation model, built from NASA satellite data, was openly available on Hugging Face.
Daniel Huynh and Jade Hardouin modified GPT-J-6B so it spread misinformation on one task while keeping benchmark scores, then uploaded that model to Hugging Face as a supply-chain demonstration.
The poster shared SummarizeThis, a free web app built with LangChain and ChatGPT that summarizes text, YouTube videos, PDFs, and tweet threads.
Rizel Scarlett wrote up the seven steps of building a first browser extension with GitHub Copilot.