I have a habit I suspect many of you share. When I find an article, a TikTok, or a LinkedIn post I want to come back to, I email it to myself. The emails go into a folder called To Watch or Read. In theory, I return to the folder and watch or read the things. In practice, the folder had grown to 181 emails, the oldest from February, and it had crossed the line from resource to guilt pile. A backlog that big is not a reading list. It is a place where potentially useful information goes to die.
So, this morning, I asked Claude to clean it out. Not to delete anything, but to open every email, follow every link, and tell me what each item actually says. The result was a Word document organized by topic: court cases and sanctions first, then studies and reports, then tool news, then everything it could not reach. Each readable item got a short summary with the author, the date, and why it might matter to my work. Each unreadable item got a one-line entry instead of a fake summary. The whole folder took three passes over one afternoon. I actually hit my token limit about halfway through the work. When I returned to my computer a few hours later, I told Claude to continue from where it left off.
I completed this task with Claude Cowork, using Fable 5.0. I was testing Fable 5.0 before it moves to pay per token, so I thought this would be a good task. That said, I can use other models to run this same process. Fable probably just completed it a bit faster than other models.
What Was in the Folder
About a third of the items were fully readable, and some were definitely worth it. The sanctions order in Withers v. City of Aberdeen, where a federal judge in Mississippi disqualified all four lawyers in a case because both sides filed briefs with AI-fabricated citations. Coverage of the Oregon Supreme Court dismissing a petition over fake AI citations. A first-person account from an open-source maintainer who rejected an AI agent’s code contribution and then watched the agent publish a hit piece about him. Insurance survey data showing AI malpractice claims arriving in the real world. I had saved good material. Material I wanted to use in my teaching and writing. I would have lost it or had to waste time rediscovering it though, something I have had to do many times in the past.
Another third of my links were LinkedIn posts, which cannot be read by an automated tool without a login. For those, the document lists who wrote each post and the first words of it, pulled from the link itself, so I can access them in my browser rapidly if they are still worth it instead of clicking though slowly with no context.
The last third were TikToks. Unfortunately, TikTok share links expire after roughly three months. Nearly every video I saved in February was gone. Not the video itself, but my path back to it. The only things that survived were the creator’s name and, where I had bothered to type one, my own note about why I saved it. “Perplexity claiming ownership of content in new terms.” “Agentic drift.” Those notes turned out to be the most reliable part of my terrible system.
One particularly useful thing that Cowork caught is a deadline I put in the subject line. The deadline was for tomorrow because it involved trying something with Fable beforce it goes to per token pay.
Two Lessons Before The AI Part
First, if you save videos by sharing links to yourself, add a note identifying what the video is about and why you care. When the link dies, and it often will, the note remains. I frequently come across these videos while I am scrolling TikTok in the evening. I don’t want to watch them at 11 at night, but they often have useful AI tips. It will be much easier for me to find the original TikToks when I noted what the video was about.
Second, process the backlog while it is fresh. The difference between running this in March and running it in July was about sixty dead links. Will I take the time to try to find the data? Probably not. I didn’t take the time to watch the videos while they were fresh, so it isn’t likely I will get myself to find the information now. I didn’t lose anything important, just tips and tricks. But obviously, at one point, I thought they were important enough to look into later.
Turning the Process Into a Skill
Doing this once was useful. Doing it every month is a habit I need to create so I turned the whole process into a skill. A skill, if you have not used one, is an instruction file that AI, in this case Claude, reads so it knows how to perform an action. For example, Claude has a skill that tells it how to make a Word document or a PDF. Skills are useful for tasks you do over and over again the same exact way.
My skill for reviewing saved articles and videos is specific to me and my mailbox, so I am not sharing the file, but I will tell you what it contains. If you tell your AI agent that you want to do something similar to what I did, the below information will be helpful to it.
- Where my folder is located and how to read it, including the specific quirks of the email connector, which has a few. My AI assistant should not rediscover known bugs every run. That is just a waste of time and tokens.
- What to do when the folder looks empty because the Outlook rule that files my saves has not run yet. The solution is to check my recent self-addressed mail.
- Fetching rules by platform. TikTok can only ever be summarized from its caption, never from guessed video content. LinkedIn is login-walled, so record the author and move on. This lets me decide if the thing I sent myself was really worth reading before I go and look. Blocked sites get marked inaccessible; the AI doesn’t waste time trying to find ways around things it cannot ever access.
- An honest status for every item: summarized, caption only, link live but caption unavailable, inaccessible, or no link found. The AI does not provide padding. This minimizes invented content, but remember, AI hallucinates.
- My document format and writing rules, so the digest reads like me and not like a press release. This comes both from my standing orders and my initial prompt instructions.
- A standing instruction to flag anything time-sensitive at the top. For example, my note to act by July 7th told the AI that this was something that needed to be addressed immediately. I didn’t tell Claude to note this, but it did it on its own. You might add this kind of alert specifically to your prompt the first time, before you create the skill.
- Last, because I am a lawyer: every case summary carries a reminder to verify details against the actual opinion before citing anything. AI summaries can be wrong or completely hallucinated. Any number of lawyers who have been bench slapped and sanctioned due to hallucinations can tell you what skipping verification costs. I will still verify every piece of information myself before I ever rely on it or share it.
Remember Confidentiality
One important note for the lawyers reading this. My folder was just links that I emailed myself, nothing confidential. Before you point any AI tool at a mailbox, think about what else is in there. Make sure it is something that can be shared with AI and that the AI is appropriately confidential. Please remember that confidentiality obligations do not pause because the tool is convenient.
An Empty Folder
The folder is empty now. The backlog is a fifteen-page document I can read and search. New saves will get processed monthly automatically using Claude Code. This means the links will still be alive and useful instead of lost.
I recognize that my habit of hoarding things I want to read later has not changed. I have had this habit for a very long time. In the past, I frequently never got around to reading what I had saved. Now, with AI, a monthly skill run will make sure the things I thought were important enough to watch or read won’t get wasted. And that’s a good thing.