Every model has a training cutoff. The software you use every day does not. That gap is why your AI describes a menu that moved, a feature that shipped, or a price that changed, and sounds certain doing it. The fix is not a better prompt. It is a reference folder.
Ask about a feature released after the cutoff and most models will not stop to say so. They answer from the nearest thing they saw in training, which is usually last year's version of the same product. The answer arrives fluent, formatted and wrong, which is the hardest kind to catch.
You ask where a setting lives. It gives you a path that was accurate eighteen months ago. You go looking, do not find it, and assume you are the problem. Nothing in the answer signalled uncertainty, so nothing prompted you to check.
You hand it the vendor's dated release notes first. Now it answers from that, and when something is missing it can tell you so, because it has a defined edge to bump into. You get a real answer or a real gap. Both are useful.
Do this once per tool you actually depend on. Not every app you have logged into. The two or three where being wrong costs you something.
The vendor's own release notes or changelog. Not a blog roundup, not a video summary, not a forum thread. The dated page the company publishes itself.
Release notes first, because they carry dates. Then the limits pages: privacy, licensing, what the tool will and will not touch. Skip the marketing pages entirely.
In Claude, add the pages to a project or turn them into a skill. Anywhere else, paste them at the top of the thread before you ask your first question.
Date-stamp the file. A stale reference folder is worse than none, because it makes the model confident about features that already changed.
Choose what you actually use. The sources, the paste-ready instruction and the downloadable skill all update to match.
Use this in any AI tool. Open the source, copy the release notes, and drop them where the block says. The instruction is what makes the model treat your notes as authority rather than as one more opinion.
A skill loads itself whenever the topic comes up, in any chat, so you are not re-pasting. Download the file, then add it to Claude as a skill. Paste your pulled notes into the Current notes section before you use it, and replace that section on every re-pull.
Every link below is the vendor's own dated page, checked on 6 August 2026. Bookmark the ones you use. These are the pages to re-pull from, not the ones to read once.
Dated updates for the Claude apps, Cowork and integrations. Start here if you use Claude to think.
support.claude.com › release notesAPI, SDKs and console changes. Has an RSS feed, so you can watch it instead of remembering to check.
platform.claude.com › release notesWhat actually shipped to general availability, split by platform. The single best page for Copilot currency.
learn.microsoft.com › release notesPrivacy, data boundary, governance, licensing. Written for IT, which is exactly why the limits are stated plainly.
learn.microsoft.com › Microsoft 365 CopilotDated feature changes for ChatGPT across plans. Separate pages exist for Enterprise, the desktop apps and agents.
help.openai.com › ChatGPT release notesWeekly recaps covering Gmail, Docs, Drive, Meet and Gemini in Workspace. Pull the recaps, not the product pages.
workspaceupdates.googleblog.comA reference folder is a small amount of maintenance that buys a large amount of accuracy. It stops buying anything the moment you stop maintaining it.
Put the pull date in the file itself, not just the filename. When the model quotes it back, the date comes with it, and you can see how much to trust it.
Swap the old notes out. Stacking three versions of the same release notes teaches the model to blend them, which is how you get features that never existed.
The instruction that earns its place is the one telling the model to flag gaps instead of filling them. An AI that admits an edge is more useful than one that never has one.
Because it has a training cutoff and the product does not. Software ships continuously. When a model is asked about something released after its cutoff, it rarely says it does not know. It answers from the closest thing it saw in training, which is often last year's version of the same product.
No. Search is per-question and depends on what the model chooses to look up in that moment, and on which page happens to rank. A reference folder is standing context you chose: the same vetted, dated pages every time. Use both. Search fills gaps, the folder sets the baseline.
Monthly for most tools, and right away after a major release. If that sounds like too much, watch an RSS feed where the vendor offers one and refresh when something lands. The Claude Platform notes publish a feed.
A project holds reference files for the conversations inside it. A skill is a portable instruction file that loads itself whenever the topic comes up, in any chat. Use a project when the work is contained. Use a skill when you want the knowledge to follow you everywhere.
It can if you paste an entire documentation site, which is why the steps say release notes first. Notes are short, dated and dense with exactly what changed. If a page is enormous, keep the last two or three months and drop the rest.
A reference folder fixes what your AI knows. It does not decide what you are trying to accomplish. That part is the strategy, and it is the part Prickled builds with you.
Start a build →