Claude's Record-a-Skill cut my research from hours to 30 minutes – but the magic has limits

Follow ZDNET: Add us as a preferred source on Google.

ZDNET's key takeaways

• Claude turns narrated screen recordings into reusable skills.
• My skill cut hours of tedious research to just 30 minutes.
• The tradeoffs: high usage, slow runs, and fragile control.

Last month, Anthropic announced a new Claude Cowork feature: the ability to narrate a screen recording and turn it into a Claude skill. I tried it and, well, as they say, "All magic comes with a price."

Let's set the stage by defining what we're after here. A skill is a very detailed canned prompt that can be run like a program from within Claude (and many other AIs).

Also: This new Claude skill saves you from bad contracts – and costs less than a lawyer

The best example is this skill from the Claude Small Business plugin that comprehensively reads and evaluates contracts. It doesn't perform as well as the best contract lawyers out there, but it's actually quite a bit more helpful and comprehensive than some of the mediocre attorneys I've worked with over the years.

Another feature of Claude Cowork is the ability for Claude to control your computer, move between your apps, browse web pages, and make decisions during that process. Think of this as the ability to run very powerful batch scripts or shortcuts.

Over the years, I've attempted to use all sorts of tools to accomplish computer automation. I've had computer automations reading pixels on the screen, attempting very complex if/then branches, and working their way down a ton of very fragile steps. Most automations take days to set up and then don't survive past a single UI change on the part of whatever it is you're automating. It will be interesting to see if this also holds true with AI automations or if they'll be able to dynamically adapt to minor UI changes on their own.

Also: I use Anthropic's Claude AI tools for very different jobs: How to pick between models, Code, and Cowork

The new Record-a-Skill feature makes automations for you. In one evening, I created a very complex and almost wildly powerful skill. The process was far easier than doing complex scripting, but it was also fussy and fiddly and required a bunch of tweaks.

In this article, I'll walk you through one and show you the magic. And then, at the end, we'll discuss the price that comes with this magic. Read through to the end. That's the important part.

Sentiment analysis and quote aggregation

Every few months, I have to do a task that is incredibly tedious and annoying but necessary for some article I'm writing. There are two services, HARO and Qwoted, that specialize in connecting experts who want to be quoted and interviewed with journalists and analysts who might choose to quote or interview them.

Here are a couple of recent examples in which I've used these services:

You get the idea. The process is fairly simple. I create a query that describes my question and sets limits on who should respond. I always say that I want to speak with the primary responder, not a PR agent who's trying to make points by getting their client some press.

These are very valuable services, but as with the Record-a-Skill feature itself, all magic comes with a price. Placing a query on HARO or Qwoted always results in a deluge of responses, many of which are irrelevant, not good enough, unfocused, pitching their own agenda, begging for interviews, and so on.

I'm working on an article comparing Claude Code to OpenAI's Codex and asked for developers to share their preferences. My deadline for responses ends in two days, but so far, I've gotten 121 replies.

(Disclosure: Ziff Davis, ZDNET's parent company, filed an April 2025 lawsuit against OpenAI, alleging it infringed Ziff Davis copyrights in training and operating its AI systems.)

Also: I used Claude Code to vibe code a Mac app in 8 hours, but it was more work than magic

The process of sifting through the replies, choosing the best ones, filtering out the noise, tabulating any data from them, and otherwise preparing them for possible use in an article usually takes me an entire day. A very clerical, very tedious, very annoying day where, by the end, my tolerance for my fellow humans declines to almost zero.

The complexity is not made any easier by the services. HARO sends emails with the answers and also makes the answers available on its site. Qwoted answers can only be seen through its dashboard. Answers aren't fully displayed, so you have to expand each answer. Contact information isn't always fully represented, so you have to click through profiles to get at who's who.

The actual answers themselves are buried in piles of self-serving intro statements, positioning pitches, closing offers, and tons of PR folks claiming their client is special. Every query I've ever made usually includes about 30% that read like this (an actual response): "I know you said no PR people, but I promise, I'm just the messenger."

Grrrr…

If there's any project suitable for AI analysis, this is it.

Getting started

To record a skill, you have to be in the Claude Desktop application. You also have to be on the Chat/Cowork side (meaning not in Claude Code mode). You have to make sure the Cowork tab (not the Chat tab) is selected. And then tap the + and choose Record a skill from the menu.

You should also make sure you have a microphone hooked up and that the computer can hear you clearly. You're going to be verbally instructing Claude along with demonstrating what you want done. So make sure your mic is good enough.

Also: How to install and configure Claude Code, step by step

Once you choose Record a skill, you'll be presented with the Start recording dialog. Go ahead and hit Start recording, and you'll be off to the races.

Get into the teaching mindset

If you've ever recorded a macro by starting a recording mode and then clicking each action in order to have the computer ultimately replace it — this is not that. Yes, the skill will replay some of the clicks you've done in your recording session, but the macro mindset isn't the right place to be.

Instead, put yourself in a teaching mindset. Assume you're showing another person how to carry out a set of tasks. Think of it as training an assistant about a complex workflow, not an algorithm to repeat clicks.

Later, once you read back the skill Claude creates, you'll see why this is so important. This process is very much teaching, not programming.

Also: I tried a Claude Code rival that's local, open source, and completely free – how it went

For my project, once I hit record, I started by explaining verbally what the project was meant to accomplish. I did this in exactly the same way I would have explained to an assistant that I was aggregating information from various submitters as input for an article I'm writing. I explained that I get a ton of responses and that many of them are unusable as quotes but still usable as data. So as we filter through all the responses (and we need to go through all the responses), we're looking for both data value and quote value.

I started already logged into HARO and Qwoted. The very next thing I did was jump to those tabs and explain what each service did, how to find the latest query, and how to interpret the services' user interfaces. As I was doing that, I was explaining how to think about what I'm looking for and then how I want it presented.

All of this explaining took me a full 10 minutes, at which point I was yanked out of the Record a skill mode and slammed back into Claude.

It's OK. I didn't expect it the first time, but it's OK.

Claude explained to me that it records in 10-minute chunks. If it hits 10 minutes, it stops recording and actually starts trying to build the skill. You can interrupt that and begin a new recording. The key is that as soon as you begin that recording, tell it the context. In my case, it was, "this continues the HARO and Qwoted sentiment report workflow."

I wound up recording three blocks, two of 10 minutes each and another of about five minutes. As I spoke, I navigated around the various listings dashboards and explained to Claude what I was paying attention to, what I wanted to prioritize, and why.

Also: How to learn Claude Code for free with Anthropic's AI courses – one took me just 20 minutes

I also opened up a Google Doc and showed Claude how I moved information from the various dashboards into the Doc and then how I wanted the Doc to be updated to give me overall aggregate information as well as detailed quotes.

Tweaking the skill

When I was done, I hit stop, and Claude presented me with the full text of the skill.

I didn't like how the skill instructions implied that whenever I made mention of HARO, it should launch, so I instructed it to change that text.

Also: How Claude Code's new auto mode prevents AI coding disasters – without slowing you down

As it turns out, you can tweak a skill after it's recorded simply by telling Claude what you want. You can also update the skill even after it's been installed. I found this very powerful and was able to fine-tune the skill for later use.

Once it's all done, save the skill.

Here are two documents. The SKILL.md file is a Markdown file that's displayed in the right pane of the Claude interface. That lets you read it to make changes. The actual skill is haro-qwoted-sentiment-report. When that's saved, it's actually installed in Claude.

Invoking the skill

Invoking the skill is easy. Just hit the slash (/) key and then type the name of the skill you want to run.

When I ran the skill, it took about half an hour to complete. When it was done, it provided me with a link to a Google Doc. Here you can see the beginning of that Doc.

Don't take these numbers specifically because they will be updated for the final article that uses this research, but you can see the kind of data aggregation that this tool is providing.

All magic comes with a price

This feature is no different. I tried using Record-a-Skill on a number of projects, and I found there were four aspects where the magic did come with a price.

Also: I used Claude Code to vibe code an Apple Watch app in just 12 hours – instead of 2 months

1. Consumes a lot of usage: The first is that the recording process, as well as the skill executions, seems to use a lot of Claude's usage limits. I have the $100-per-month Max plan. A few hours of Claude Cowork usage controlling my computer used an entire session's worth of usage allocation. I got dropped into the billable credits interface. What's so odd is that I have spent entire days in Claude Code and never triggered the same usage overage.

2. Computer use is very slow: The computer use capability is helpful but incredibly slow. It's almost agonizing to wait for Claude to move from one interface element to another. This is fine when you start a skill running and leave for dinner. But if you want a response quickly, forget it.

3. Computer use can dominate your machine: While you can get some work done during aspects of skill operation, if the skill is trying to navigate actual locations in windows on your screen, anything you do will cause it to make mistakes, screwing up your skill's run.

4. Claude doesn't always notice computer mistakes: If Claude clicks the wrong thing or you move the mouse when it doesn't expect it, sometimes Claude doesn't notice it. Claude can sit there stuck forever until you stop the skill and start it all over again. During that time, Claude is still consuming usage.

It can still be incredibly useful for the right task

Overall, I am really impressed with the quality of the skill that this process created. I am also really quite impressed with what Claude Cowork is able to do based on a single skill. Doing this kind of aggregation in the past has taken me four to eight hours of really tedious work. Creating the skill itself took me about an hour end-to-end.

Also: I recreated OpenAI's Codex Micro for a lot less – and mine is more customizable

Running the skill was a matter of just launching it, going on and doing other things, and coming back 30 minutes later. The time savings are phenomenal. The quality is, frankly, better than what I did by hand.

The key to this process is identifying tasks that lend themselves to this kind of skill development. There is a difference between what you can do with Claude Code when building code and how you can use Claude Cowork to handle tedious computer automation processes and data aggregation processes with these skills.

My advice is to carefully think through what it is you want to accomplish. If you think it lends itself to a single skill or a set of skills, then use this tool. My other piece of advice is to not be afraid to have it modify the skill once it's created. Just tell Claude how you want the skill modified or what you want it to do, and Claude will do that. I had this skill modified three or four times by the time I was done. The result is really quite powerful.

What kind of skills do you think you would like to create with this new ability? Let us know in the comments below.

You can follow my day-to-day project updates on social media. Be sure to subscribe to my weekly update newsletter, and follow me on Twitter/X at @DavidGewirtz, on Facebook at Facebook.com/DavidGewirtz, on Instagram at Instagram.com/DavidGewirtz, on Bluesky at @DavidGewirtz.com, and on YouTube at YouTube.com/DavidGewirtzTV.