I'm using Playwright so much right now. All of the good LLMs appear to know the API really well.
Using Claude Code I'll often prompt something like this:
"Start a python -m http.server on port 8003 and then use Playwright Python to exercise this UI, there's a console error when you click the button, click it and then read that error and then fix it and demonstrate the fix"
This works really well even without adding an extra skill.
I think one of the hardest parts of skill development is figuring out what to put in the skill that produces better results than the model acting alone.
Have you tried iteratively testing the skill - building it up part by part and testing along the way to see if the different sections genuinely help improve the model's performance?
this is the core problem rn with developing anything that uses an LLM. It’s hard to evaluate how well it works and nearly impossible to evaluate how well it generalizes unless the input is constrained so tightly that you might as well not use the LLM. For this I’d probably write a bunch of test tasks and see how well it performs with and without the skill. But the tough part here is that in certain codebases it might not need the skill. The whole environment is an implicit input for coding agents. In my main codebase right now there are tons of playwright specs that Claude does a great job copying / improving without any special information.
edit with one more thought:
In many ways this mirrors building/adopting dev tooling to help your (human) junior engineers, and that still feels like the good metaphor for working with coding agents. It's extremely context dependent and murky to evaluate whether a new tool is effective -- you usually just have to try it out.
Yeah you can definitely do this with prompts since LLMs know the API really well. I just got tired of retyping the same instructions and wanted to try out the new Skills.
I did test by comparing transcripts across sessions to refine the workflow. As I'm running into new things I'm continuing to do that.
When it works, its totally magic, but I find it gets hung up on things like not finding the active playwright window or being able to identify elements on the screen.
One thing that I see skills having the advantage is when they include scripts for specific tasks that the LLM has a difficult time generating the right code.
Also the problem with the LLM being trained to use foo tool 1.0 and now foo tool is on version 2.0.
The nice thing is that scripts on a skill are not included in the context and also they are deterministic.
I was looking into creating one and skimmed the available ones and didn't see it.
EDIT:
Just looked again. In the docs they have this section:
```
Available Skills
Pre-built Agent Skills
The following pre-built Agent Skills are available for immediate use:
PowerPoint (pptx): Create presentations, edit slides, analyze presentation content
Excel (xlsx): Create spreadsheets, analyze data, generate reports with charts
Word (docx): Create documents, edit content, format text
PDF (pdf): Generate formatted PDF documents and reports
These Skills are available on the Claude API and claude.ai. See the quickstart tutorial to start using them in the API.
```
I have created a simple .sh command to do the testing using browser-use
and add how to prompt this command in CLAUDE.md
because it has to run a shell command, it is more deterministic than Skills.
And it uses near to zero of context window compared to MCP.
Recently, I have found myself getting more interested in shell commands than MCPs. There is no need to set it up. Debugging is far easier. And I would be free to use whichever model I like ot use for a specific function.
For example, for Playwright, I use GPT-5 just because I have free credits. I could save my Claude Code Quota for more important tasks.
I have a few questions about test frameworks that use AI services like this.
1)The examples always seem very generic: "Test Login Functionality, check if search works, etc". Do these actually work well at all once you step outside of the basic smoketest use cases?
2) How to you prevent proprietary data from being read when you are just foisting snapshots over to the AI provider? There's no way I'd be able to use this in any kind of real application where data privacy is a constraint.
1) Beyond basic tests: You're right to be skeptical. This is best for quick exploratory testing during local development ("does my new feature work?"), not replacing your test suite. Think "scriptable manual testing" - faster than writing Playwright manually, but not suitable for comprehensive CI/CD test coverage.
2) Data privacy: Screenshots stay local in /tmp, but console output and page content Claude writes tests against are sent to Anthropic. This is a local dev tool for testing with dummy data, not for production environments with real user data. Same privacy model as any AI coding assistant - if you wouldn't show Claude your production database, don't test against it with this.
Thanks. I keep seeing silver bullet testing solutions pitched left right and center and wondering about these two points. Glad to see a project with realistic boundaries and expectations set. Would definitely give this a shot if I was working on a local project.
For 2, a lot of companies use AWS Bedrock to access Claude models instead of Anthropic, for exactly this reason. Amazon’s terms say they don’t log prompts or completions and don’t send anything to the model provider. If your production database is already hosted by AWS, it doesn’t seem like much additional risk.
> Do these actually work well at all once you step outside of the basic smoketest use cases?
Excellent question... no, beyond basic kindergarten stuff playwright (with AI) falls quickly apart. Have some OAuth? Good luck configuring playwright for your exact setup. Need to synthesize all information available from logs and visuals to debug something? Good luck..
but how would claude then look at devtools for the playwright window to see console output? i know some frameworks are putting logging into the shell, but in old repos Playwright MCP seems worth the extra context, no?
If you're willing to have Claude write code to test a thing you could do a teeny bit more work and make that Playwright script a permanent part of your codebase. Then, the script can run in your CI on every build, and you can keep enhancing it as your product changes so it keeps proving that area of your product works as desired.
Have it run inside a harness that spins up its own server & stack (DB etc.) and boom - you now have an end-to-end test suite!
How do images affect context? Does an image run separately on another model and returns a text description of it that ends up smaller than the accessibility text tree?
Heh BrowserBase and Browser-Use exist specifically because this is a harder problem than it looks. Any approach will work for the first couple actions, that hard parts are long strings of actions that depend on the results of previous actions, compressing the context and knowing what to send, and having your tools work across all the edge cases (e.g. date picker fields, file upload fields, cross origin iframes, etc.).
Related anecdote: some months ago I tried to coax the Playwright MCP to do a full page screenshot and it couldn't do it. Then I just told Claude Code to write a Playwright JS script to do that and it worked at the first try.
Taking into account all the tools crap that the Playwright MCP puts in your context window and the final result I think this is the way to go.
Working on this problem but a combo of a skill and an mcp better suited to playwright is the solution IMO.
The issue is for many things playwright is really verbose, by better tailoring outputs and making them more fine grained you’ll get less context bloat and allow the llm to better work with the context. I’m making it open source.
MCPs themselves may provide access to tools that are either deterministic or not, but the LLM using them generally isn't deterministic, so when used by an LLM as part of the request-response cycle determinism, if the MCP-provided tool had it, is not in a feature of the overall system.
SKILLS.md relies on a deterministic code execution environment, but has the same issue. I'm not seeing a broad difference in kind here when used in the context of an LLM response generation cycle, and that’s really the only context where both are usable (MCP could be used for non-LLM integration, but that doesn't seem relevant.)
Yeah, this isn’t meant to replace your real tests it’s more for quick “does my new feature work?” checks during local dev. Think of it like scriptable manual testing: Claude spits out the Playwright code faster than you would, but it’s not CI-level coverage.
And for privacy screenshots stay local in /tmp, but console output and page content do go to Claude/Anthropic. It’s designed for dev environments with dummy data, not prod. Same deal as using Claude for any coding help.
If you're going to use claude to help you respond to feedback the least you can do is restate this in your own words. Parent commenter deserves the respect of corresponding with a real human being.
Yes, they are. Like any standard API, it is an orchestration layer that, given a specific input should always execute the same logic and produce a consistent output. Its job is deterministic execution.
The agentic system that uses MCP (e.g., an LLM) is fundamentally non-deterministic. The LLM's decision of which tool to call, when to call it, and what to do with the response is stochastic.
I think that this is actually the biggest threat to the current "AI bubble." Model efficiency and diffusion of models to open source. It's probably to start hedging bets on Nvidia
Most of the big OSS AI codebases (LLM and Diffusion, at least) have code to work on any GPU, not just nVidia GPUs, now. There's a slight performance benefit to sticking with nVidia, but once you need to split work across multiple GPUs, you can do a cost-benefit analysis and decide that, say, 12 AMD GPUs is faster than 8 nVidia GPUs and cheaper, as well.
Then nVidia's moat begins to shrink because they need to offer their GPUs at a somewhat reduced price to try to keep their majority share.
Share can go up and down if consumption keeps going up crazily. We now spend more per dev on their personal use inferencing providers than their home devices, so inferencing chips are effectively their new personal computers...
I'm using Playwright so much right now. All of the good LLMs appear to know the API really well.
Using Claude Code I'll often prompt something like this:
"Start a python -m http.server on port 8003 and then use Playwright Python to exercise this UI, there's a console error when you click the button, click it and then read that error and then fix it and demonstrate the fix"
This works really well even without adding an extra skill.
I think one of the hardest parts of skill development is figuring out what to put in the skill that produces better results than the model acting alone.
Have you tried iteratively testing the skill - building it up part by part and testing along the way to see if the different sections genuinely help improve the model's performance?
this is the core problem rn with developing anything that uses an LLM. It’s hard to evaluate how well it works and nearly impossible to evaluate how well it generalizes unless the input is constrained so tightly that you might as well not use the LLM. For this I’d probably write a bunch of test tasks and see how well it performs with and without the skill. But the tough part here is that in certain codebases it might not need the skill. The whole environment is an implicit input for coding agents. In my main codebase right now there are tons of playwright specs that Claude does a great job copying / improving without any special information.
edit with one more thought: In many ways this mirrors building/adopting dev tooling to help your (human) junior engineers, and that still feels like the good metaphor for working with coding agents. It's extremely context dependent and murky to evaluate whether a new tool is effective -- you usually just have to try it out.
Also, if you figure out a good prompt today you don't know how long it will last, because of model updates outside your control
there's a console error when you click the button
Chrome Devtools also has an MCP server that you can connect an LLM to, and it's really good for debugging frontend issues like that.
Yeah you can definitely do this with prompts since LLMs know the API really well. I just got tired of retyping the same instructions and wanted to try out the new Skills.
I did test by comparing transcripts across sessions to refine the workflow. As I'm running into new things I'm continuing to do that.
I get so many LLM death spirals with playwright.
When it works, its totally magic, but I find it gets hung up on things like not finding the active playwright window or being able to identify elements on the screen.
One thing that I see skills having the advantage is when they include scripts for specific tasks that the LLM has a difficult time generating the right code.
Also the problem with the LLM being trained to use foo tool 1.0 and now foo tool is on version 2.0.
The nice thing is that scripts on a skill are not included in the context and also they are deterministic.
I'm surprised Anthropic didn't release skills with a `skill-creation` skill.
But they did.
Did they!? Damn I missed it.
I was looking into creating one and skimmed the available ones and didn't see it.
EDIT:
Just looked again. In the docs they have this section: ``` Available Skills
Pre-built Agent Skills The following pre-built Agent Skills are available for immediate use:
These Skills are available on the Claude API and claude.ai. See the quickstart tutorial to start using them in the API. ```Is there another list of available skills?
Their repo here: https://github.com/anthropics/skills
This is the skill creation one: https://github.com/anthropics/skills/blob/main/skill-creator...
You can turn on additional skills in the Claude UI from this page: https://claude.ai/settings/capabilities
Nice, thanks!
I have created a simple .sh command to do the testing using browser-use and add how to prompt this command in CLAUDE.md because it has to run a shell command, it is more deterministic than Skills. And it uses near to zero of context window compared to MCP.
Recently, I have found myself getting more interested in shell commands than MCPs. There is no need to set it up. Debugging is far easier. And I would be free to use whichever model I like ot use for a specific function. For example, for Playwright, I use GPT-5 just because I have free credits. I could save my Claude Code Quota for more important tasks.
I have a few questions about test frameworks that use AI services like this.
1)The examples always seem very generic: "Test Login Functionality, check if search works, etc". Do these actually work well at all once you step outside of the basic smoketest use cases?
2) How to you prevent proprietary data from being read when you are just foisting snapshots over to the AI provider? There's no way I'd be able to use this in any kind of real application where data privacy is a constraint.
Good questions!
1) Beyond basic tests: You're right to be skeptical. This is best for quick exploratory testing during local development ("does my new feature work?"), not replacing your test suite. Think "scriptable manual testing" - faster than writing Playwright manually, but not suitable for comprehensive CI/CD test coverage.
2) Data privacy: Screenshots stay local in /tmp, but console output and page content Claude writes tests against are sent to Anthropic. This is a local dev tool for testing with dummy data, not for production environments with real user data. Same privacy model as any AI coding assistant - if you wouldn't show Claude your production database, don't test against it with this.
Thanks. I keep seeing silver bullet testing solutions pitched left right and center and wondering about these two points. Glad to see a project with realistic boundaries and expectations set. Would definitely give this a shot if I was working on a local project.
Another llm-generated response. This is sad.
For 2, a lot of companies use AWS Bedrock to access Claude models instead of Anthropic, for exactly this reason. Amazon’s terms say they don’t log prompts or completions and don’t send anything to the model provider. If your production database is already hosted by AWS, it doesn’t seem like much additional risk.
> Do these actually work well at all once you step outside of the basic smoketest use cases?
Excellent question... no, beyond basic kindergarten stuff playwright (with AI) falls quickly apart. Have some OAuth? Good luck configuring playwright for your exact setup. Need to synthesize all information available from logs and visuals to debug something? Good luck..
Why not just ask the agent to use Playwright via CLI? That’s what I do and it works fine. With Codex anyway
Edit: oops that’s what you did too. Yes most MCP shouldn’t be used.
but how would claude then look at devtools for the playwright window to see console output? i know some frameworks are putting logging into the shell, but in old repos Playwright MCP seems worth the extra context, no?
This is brilliant.
If you're willing to have Claude write code to test a thing you could do a teeny bit more work and make that Playwright script a permanent part of your codebase. Then, the script can run in your CI on every build, and you can keep enhancing it as your product changes so it keeps proving that area of your product works as desired.
Have it run inside a harness that spins up its own server & stack (DB etc.) and boom - you now have an end-to-end test suite!
How do images affect context? Does an image run separately on another model and returns a text description of it that ends up smaller than the accessibility text tree?
Heh BrowserBase and Browser-Use exist specifically because this is a harder problem than it looks. Any approach will work for the first couple actions, that hard parts are long strings of actions that depend on the results of previous actions, compressing the context and knowing what to send, and having your tools work across all the edge cases (e.g. date picker fields, file upload fields, cross origin iframes, etc.).
How does this compare to using chrome-devtools mcp? I've been using that for validating UI flows and I haven't hit limits with Claude yet
Thanks, I have installed it and it works great!
Related anecdote: some months ago I tried to coax the Playwright MCP to do a full page screenshot and it couldn't do it. Then I just told Claude Code to write a Playwright JS script to do that and it worked at the first try.
Taking into account all the tools crap that the Playwright MCP puts in your context window and the final result I think this is the way to go.
Working on this problem but a combo of a skill and an mcp better suited to playwright is the solution IMO.
The issue is for many things playwright is really verbose, by better tailoring outputs and making them more fine grained you’ll get less context bloat and allow the llm to better work with the context. I’m making it open source.
This was on my TODO list for the week, thanks for sharing!
Now I just need to make a skill for using Jira and I can go back to the MCP-free life.
thanks!
How does this compare to using the recently released chrome mcp?
MCPs are deterministic, SKILLS.md isn't. Also run.js can run arbitrarily generated Node.js code. It is a trivial vector for command injection.
This might be sufficient for an independent contractor or student. It shouldn't be used in a production agent.
> MCPs are deterministic, SKILLS.md isn't.
MCPs themselves may provide access to tools that are either deterministic or not, but the LLM using them generally isn't deterministic, so when used by an LLM as part of the request-response cycle determinism, if the MCP-provided tool had it, is not in a feature of the overall system.
SKILLS.md relies on a deterministic code execution environment, but has the same issue. I'm not seeing a broad difference in kind here when used in the context of an LLM response generation cycle, and that’s really the only context where both are usable (MCP could be used for non-LLM integration, but that doesn't seem relevant.)
LLMs are not deterministic though. So by definition MCPs are not deterministic.
For example, GPT-5 doesn't support temperature parameter. And even models that do support temperature are not deterministic with temperature=0.
Yeah, this isn’t meant to replace your real tests it’s more for quick “does my new feature work?” checks during local dev. Think of it like scriptable manual testing: Claude spits out the Playwright code faster than you would, but it’s not CI-level coverage.
And for privacy screenshots stay local in /tmp, but console output and page content do go to Claude/Anthropic. It’s designed for dev environments with dummy data, not prod. Same deal as using Claude for any coding help.
If you're going to use claude to help you respond to feedback the least you can do is restate this in your own words. Parent commenter deserves the respect of corresponding with a real human being.
You don’t understand what you’re talking about enough so you have to ask llm to generate a response for you?
MCPs aren't deterministic...
Yes, they are. Like any standard API, it is an orchestration layer that, given a specific input should always execute the same logic and produce a consistent output. Its job is deterministic execution.
The agentic system that uses MCP (e.g., an LLM) is fundamentally non-deterministic. The LLM's decision of which tool to call, when to call it, and what to do with the response is stochastic.
i just use https://github.com/ravitemer/mcphub.nvim and enable/disable the mcps i want
I am using Tmux lynx browser MCP for sites that don't need JavaScript
I think that this is actually the biggest threat to the current "AI bubble." Model efficiency and diffusion of models to open source. It's probably to start hedging bets on Nvidia
Why would OSS models threaten Nvidia?
Most of the big OSS AI codebases (LLM and Diffusion, at least) have code to work on any GPU, not just nVidia GPUs, now. There's a slight performance benefit to sticking with nVidia, but once you need to split work across multiple GPUs, you can do a cost-benefit analysis and decide that, say, 12 AMD GPUs is faster than 8 nVidia GPUs and cheaper, as well.
Then nVidia's moat begins to shrink because they need to offer their GPUs at a somewhat reduced price to try to keep their majority share.
Share can go up and down if consumption keeps going up crazily. We now spend more per dev on their personal use inferencing providers than their home devices, so inferencing chips are effectively their new personal computers...
> There's a slight performance benefit to sticking with nVidia
In training, not in inference and not in perf/$.