How it works
SteadyVane tracks what's actually rising in AI tools, with the numbers, for people who use AI tools to get work done. It's free with no ads. Every brief is checked before it goes out.
1. It watches
Hacker News, GitHub, Hugging Face, Product Hunt, YouTube, the VS Code Marketplace and Google Trends are checked through the day, most of them every few hours. Each check writes down every item's numbers. SteadyVane has taken 457,563 readings so far (including Reddit's, which it read until June 2026).
The numbers are each source's own: YouTube views, Hacker News upvotes (Hacker News calls them points), GitHub stars (saves), Product Hunt votes, Hugging Face likes and VS Code installs. Google Trends gives a ranking score, not a count.
2. It spots what's rising
It flags items climbing much faster than usual on their source, then keeps only the ones judged to matter to people who use AI tools for work. It leaves out YouTube videos whose views look paid for and items it finds are not in English.
3. An AI writes the daily brief
Every brief has up to three stories: what happened, the numbers that prove it (from every source that carries it) and why it matters if you work with AI tools, plus a short take. A story isn't repeated within a week unless something new happens. The brief skips videos it finds with a sponsor, a promo code or an affiliate or referral link.
4. It's checked before it goes out
- Fact-check: the main claim is checked on the web. If it's false or doubtful, it isn't written up. If it can't be confirmed, it's reported only as someone's claim ("the maker says").
- Numbers: code checks them against SteadyVane's measurements, the source's own words and the facts the fact-check confirmed. A number it can't find there goes back to the writer once; if the writer keeps it, the story doesn't go out.
- Every sentence: a separate AI check reads each one against the evidence. Anything it can't back up is rewritten or cut.
- Grade: each story is scored for accuracy and usefulness. Weak ones are dropped.
- A person: gets the brief half an hour before it publishes and can hold or fix it. If that step can't happen, nothing publishes.
The Rising tracker shows measurements only; its items aren't fact-checked.
What "about 20× the typical pace" means
The item gained about 20 times as fast, per hour, as the typical rising item SteadyVane tracks on the same source over the same hours. It's shown only when there are enough items to compare with and the item is clearly faster.
Where to get it
- Here: the daily brief, the Rising tracker, every past brief and a page for each item with its numbers day by day.
- Every week: the week's top items (Tuesdays, also by email), the standings of AI coding tools (Wednesdays) and the Receipts (Fridays). The Receipts show how the items the brief featured a week or more ago are doing now (read again at their source every day), including the ones that went nowhere.
- Follow: each brief goes to Telegram, Bluesky, X and Threads. A short video of it goes out later that day there and on YouTube and Instagram.
Privacy and contact
This site sets no cookies; the host counts visits without them (see Privacy).
When something on this site turns out to be wrong, it is corrected.
Questions or corrections: hello@steadyvane.com
The exact rules
Public sources, around the clock
Hacker News (front page and search), GitHub (new repositories and a watchlist of established tools), Hugging Face trending models and spaces, Product Hunt, YouTube (searches for AI tools), the VS Code Marketplace (AI extensions) and Google Trends are checked through the day, most of them every few hours (Product Hunt twice a day). Each check records what every item measured at that moment - 457,563 observations so far (including Reddit's, read until June 2026) - so the system knows not just what is popular but what is moving.
The counts are the sources' own: a YouTube video's views, a Hacker News post's upvotes (Hacker News calls them points; it has no down arrows on stories), a GitHub repository's stars (saves: starring a repository saves it to find again later), Product Hunt votes, Hugging Face likes and VS Code Marketplace installs. A Google Trends search's number is not a count: it is a score from the search's place in Google's list of trending searches (the higher, the nearer the top). Briefs written before mid-October 2026 say "points" for Hacker News upvotes.
What counts as rising
A velocity detector flags the items whose readings are climbing much faster than usual for their source (or the part of it they come from, such as Hacker News' front page); they become candidates. YouTube videos whose views look paid for (almost no likes for their view count) are set aside. So are items the language check finds are not in English. The brief also skips YouTube videos whose description carries a sponsor, a promo code or an affiliate or referral link. A relevance reviewer scores each one for how much it matters to people who use AI tools to get work done; only items above the bar appear on the tracker or in a brief. Items about the same thing - a launch, a tool, a technique - become one story in the brief, and a story is not repeated within a week unless something new happens.
The pace
When SteadyVane says an item climbed at about 20 times the typical pace, it means this: the item's gain per hour between two of its readings (in a brief, its first and latest reading before the brief was written), divided by the median gain per hour of the other items SteadyVane read on the same source in those same hours - each over its own readings there, at least an hour apart, counted once, and only those that climbed and whose count did not fall by more than 2%. A pace is stated only when at least eight other items climbed in those hours and the item was at least 1.5 times faster. A count that went down by more than 2% between two readings never gives a pace (from mid-October 2026 such an item is not written up in a brief), and a Google Trends score never gives one.
Five gates before a brief publishes
- Fact-check. The central claim of each candidate is checked against the open web. A claim the check finds false or doubtful is never written up; one it does not confirm (also when the check itself fails) is published only as someone's claim ("the maker says").
- Receipts. The brief is written from the recorded measurements, the source's own words and the facts the fact-check confirmed. Every number in the text is checked by code against them - a number may be rounded (within 6%), and small numbers (twelve or less, such as a version) and years are let through; the text goes back to the writer once, and an item that still carries a number that is not there does not publish.
- Every sentence. A separate AI check (the same model as the writer, in its own call) reads every sentence the writer wrote - the brief, each social post, the Instagram caption and the day's title - against the evidence the code assembled for that item from SteadyVane's records. A sentence it cannot support goes back to the writer once, then is cut; an item whose headline or measured proof it cannot support does not publish.
- Grade. A separate AI grading pass (again the same model, in its own call) scores each item's brief text and its X, Bluesky and Threads posts for accuracy against the data, usefulness, specificity, and voice. Below the bar, it is dropped.
- A person. The finished brief goes to a human editor half an hour before it publishes; it can be held, corrected, or trimmed - a correction is the editor's own and is not re-checked by the gates above. If that check cannot happen, nothing publishes.