Blog
How Do You Keep Up With AI? Here's My Approach.
Most mornings I scroll through social media while I walk my dog, and most mornings something about Claude catches my eye: a new feature, a debate about agentic loops or agentic graphs, how building software with AI differs from the traditional development process, a new model, an old one being retired. I follow it down the rabbit hole. More than once, a single post has cost me half a day of research, just to work out whether the idea, as Anthropic describes it, would actually help my projects, and how soon.
It is a lot to keep up with, and it can feel overwhelming.
Every platform a business depends on now changes faster than its team can keep up with. AI is the extreme case. In the 30 days to October 14, 2026, the eight Anthropic sources I watch published 2,468 updates. Anthropic is the company behind the Claude AI models. Most of those updates do not matter to any given project. A few matter a lot, such as a model your software depends on being retired on a fixed date.
Reading all of it is not realistic, and skimming is how the important one gets missed. So I built a tool that reads everything and tells me what matters to my own work. It is called What's New for Claude, a free add-on for Claude Code, Anthropic's AI coding assistant. It is an independent project, not made by or affiliated with Anthropic.
The questions I started with
Before any code, I wrote down what the tool had to get right.
What does it cost to miss one? Far more than reading an extra alert. So the tool leans toward flagging, and some things, such as a retiring model my code uses, are always flagged.
Is it about my work, or just news? An update only matters if it touches something I use. So the tool checks my own project files for the model, package or setting an update mentions, on my machine, before it scores anything.
What would I do about it? Every alert comes with one next step, and a link to the original source so I can check it myself.
Can I trust what it read? Updates are written by other people, and text can be written to trick an AI. That question shaped the design more than any other.
What does it cost to run? A typical day takes about four AI requests on my own Claude account, and it spends nothing in the background unless I allow it.
How the AI is kept on a short leash
The AI that reads the updates works with no tools. In plain terms, it can read text and write an answer, and nothing else. It cannot open files, browse the web, run programs or change anything. If an update hid an instruction meant to trick an AI, the most it could do is nudge a score within its range or slant the next step the AI writes.
Everything that acts is plain software following fixed rules. It never changes my settings or my code, and the contents of my files never leave my machine. What it does send, through my own Claude account, is the update text, my project names and short descriptions, and the names it looked for and the file paths where it found them.
What the score means
Each update gets a score from 0 to 10 for each of my projects. The score answers one question: how soon should I deal with this, here?
Two things set it. The first is how serious the change is. A retiring model, a change that breaks existing setups or a security issue ranks above a new feature. The software turns that into a range, such as 7 to 10 for urgent items or 2 to 5.9 for ones that can wait. The second is whether the update touches something the project actually uses. If the file check does not find it, the update cannot reach 8 for that project, however serious it sounds.
The AI then picks the exact spot inside that range and writes the next step. The update's own wording can move it within the range, never out of it.
An 8 or above is a top pick, and it comes with a label. "Act now" means there is a change to look at. "Confirm handled" means the project already seems to meet the change, so I only need to check. Everything else lands under "Later," "Not used here" or "Already in your version."
One real example. On October 17, the tool checked a small sample project built on Claude Sonnet 4.5, a model that retires on November 30, 2026. It scored the update 9.3, "Act now," and pointed to the exact file and line to change, and the model to switch to. The software made that call on its own, with no AI involved.
Does it work?
In those 30 days, 2,468 updates came in. The tool reviewed all but two of them against my nine projects, and five reached a top-pick score. Four said "Act now." That is the point of the tool: I had five things to look at, not 2,468.
One of them was a model that one of my products still depended on, and it was being retired. I had not noticed. The tool flagged it, showed me where the model was used and helped me move to its replacement.
The tool is checked the same way. Since version 0.3.1, every release gets three independent reviews before it ships, one for correctness, one for security and privacy, and one whose only job is to prove the first two wrong. Then a check on my machine refuses to publish a release without a complete review record. It checks that the record exists, not how good the review was.
It still misses things. It does not yet read two of Anthropic's official sites, and that fix is planned for the next feature release.
A small project, the same work
Claude wrote the code. The decisions above were mine. The larger part of the work was deciding what the system may do, checking the evidence before trusting it, and putting a check behind each rule. That is the same discipline an ERP implementation demands, applied here across nine projects on a platform that changes every week.
I still read about Anthropic's newest features on my morning walks. The difference is that I can rest assured. My projects tend to stay in sync, and when something needs my urgent attention, the tool tells me in advance.
The code and the full technical details are on GitHub, and the AI/POC Lab has the project card. What's New for Claude is an independent project by Gabriel Philipson, not made by or affiliated with Anthropic.
Sources
1. Claude Sonnet 4.5 retirement: Anthropic, "Model deprecations." Deprecated September 30, 2026, retires November 30, 2026, replacement claude-sonnet-5-5 (checked October 21, 2026). https://platform.claude.com/docs/en/about-claude/model-deprecations
2. How the tool scores updates, what it sends and what a typical day costs: the plugin's README. https://github.com/gabrielp-creator/whats-new-claude
3. Release versions: the plugin's changelog. https://github.com/gabrielp-creator/whats-new-claude/blob/main/CHANGELOG.md
4. The 30-day counts (2,468 updates, 5 top picks across 9 projects, to October 14, 2026, measured October 21, 2026) come from the tool's own stored results, and the release review rule from the project's working records. Neither is public.
The project
See What's New for Claude in the AI/POC Lab
The project card covers what it does, how it was built, and where it stands.
View AI Projects
Leave a Comment