전사 🔗
WEBVTT
00:00:00.500 --> 00:00:05.472
The last time this show talked about coding agents, the tool was Cursor and the word was vibe coding.
00:00:05.884 --> 00:00:07.841
That was back in March 2025.
00:00:08.228 --> 00:00:12.538
Everything I've built since, including Salience, was built with a tool this show has never covered.
00:00:13.057 --> 00:00:14.496
And here's the bit I didn't expect.
00:00:14.589 --> 00:00:17.478
Some weeks I use it more for SEO than for code.
00:00:18.155 --> 00:00:35.170
Hi, it's me, Ed Dawson, back as always, and this is SEO Is Not That Hard, and by the end of this episode, you'll know what Claude Code actually is, how I use it to build software, the three ways I use it on SEO work every week, and one rule that will stop it making things up with your data.
00:00:36.106 --> 00:00:37.939
So what is Claude Code?
00:00:38.338 --> 00:00:41.707
Claude Code is an AI agent that lives in your terminal.
00:00:41.793 --> 00:00:45.237
So it's not in a chat window, it's not auto-completing an editor.
00:00:45.430 --> 00:00:49.429
It's literally within the terminal on your computer.
00:00:49.502 --> 00:00:56.253
And if you don't know what a terminal is, just go and ask Claude or ask ChatGPT,"What is a terminal?" And it'll tell you all about it.
00:00:56.253 --> 00:01:13.704
But it's basically the fundamental layer of a computer that's a bit deeper and lower down than you generally tend to work in But it's very powerful down there Basically, what you do is you open a folder on your computer, you type in what you want in plain English, and it goes and does it.
00:01:14.519 --> 00:01:24.504
It can read the files in a folder, it can run commands, it can look at the results, it can decide what to do next, and it can keep going until it get its-- gets its job done.
00:01:24.605 --> 00:01:26.292
It can create computer code.
00:01:26.715 --> 00:01:36.168
It will write little software for itself to run things, and it just works in loops until it gets your question or your problem solved and answered.
00:01:36.570 --> 00:01:42.066
And that loop, the way that it loops around is the whole difference between a chatbot and an agent.
00:01:42.448 --> 00:01:47.221
A chatbot just answers a question, turn response basis.
00:01:47.296 --> 00:01:51.760
You're used to it in-- when you use Claude or ChatGPT in their website.
00:01:52.727 --> 00:01:57.206
But an agent has agency, it can act, it can do things.
00:01:58.087 --> 00:02:02.537
Now, Claude Code came out as a preview in February twenty twenty-five, and properly in May.
00:02:02.834 --> 00:02:10.042
Now, I didn't properly pick it up and start using it until Christmas when I was back ill on the sofa where we spoke about it in a couple of episodes ago.
00:02:10.542 --> 00:02:11.735
And I told you where that led.
00:02:12.358 --> 00:02:15.334
We got to Salient from that just from using this tool.
00:02:16.330 --> 00:02:19.579
But here's what matters for the rest of the episode.
00:02:20.495 --> 00:02:27.951
It doesn't care whether the folder holds code or a pile of spreadsheets, which is where the SEO bit starts.
00:02:27.962 --> 00:02:30.936
So first, we'll cover the coding bit briefly.
00:02:31.436 --> 00:02:36.324
And this is just a recap really from the episode of the day and what a software factory is.
00:02:36.475 --> 00:02:37.516
So I'll keep this short.
00:02:37.871 --> 00:02:39.396
So I use it to code.
00:02:39.466 --> 00:03:00.306
I write a brief, which is the problem, an example of what done looks like, what tests must pass, and then Claude Code uses agents within itself, a manager agent that will turn that into a specification, an implementer agent which will build it And then a reviewer agent that will check the work and send corrections back.
00:03:00.665 --> 00:03:03.427
And then when it's done and it's ready, it will present it to me.
00:03:03.667 --> 00:03:04.497
I can review...
00:03:04.532 --> 00:03:05.425
I will review it.
00:03:05.605 --> 00:03:09.812
Always have a human review it, and if we're happy with it, then we'll push it out to production.
00:03:10.278 --> 00:03:16.346
That's our software factory, and that's the kind of powerful thing you can do with Claude Code.
00:03:16.764 --> 00:03:18.354
But it doesn't have to be that complicated.
00:03:18.370 --> 00:03:20.925
You don't have to be building a software factory to get stuff done.
00:03:21.197 --> 00:03:26.603
You can actually start much simpler, and that's where it actually comes in really useful for SEO.
00:03:26.618 --> 00:03:36.694
So that's what we'll talk about next So as I've mentioned before, over many episodes, I run a handful of content sites that earn through affiliate links.
00:03:36.996 --> 00:03:43.818
So for years, answering a question about them and how they were doing, how they're performing really meant sort of three tabs.
00:03:43.822 --> 00:03:49.568
That was Search Console, Analytics, and the affiliate network.
00:03:50.143 --> 00:04:04.379
And that was always exporting, pasting, pivoting, pulling and collating together by hand, and that took time, which means the more time something takes, the less likely you are to do it, especially if you're me.
00:04:04.879 --> 00:04:07.906
But this is the kind of problem that is perfect for Claude Code.
00:04:08.116 --> 00:04:18.086
So what I had Claude Code do was build me a small tool that I called WILF WOLF stands for Workspace Intelligence Layer and Facts.
00:04:18.173 --> 00:04:22.786
And I'll be honest, I came up with the name first and then worked an, an acronym back towards it.
00:04:23.038 --> 00:04:23.884
I just like the name WOLF.
00:04:23.904 --> 00:04:43.343
I wanted to be able to call it something Now what Claude Code built Wolf to do is that every night it pulls search console data, Google Analytics data from the affiliate networks that I work with, and Clicky, which is also another of, tracking system that I use that's similar but simpler than Google Analytics, but does essentially
Load full transcript
the same thing.
00:04:43.345 --> 00:04:45.180
But it's like a second tracking system.
00:04:45.180 --> 00:04:58.384
I've talked about it before in previous episodes, and I do recommend it for simple analytics And it saves all that into a local database on my Mac, and then exposes that database to Claude Code through an MCP server.
00:04:58.769 --> 00:05:03.307
So if MCP is new to you, go back and listen to the May twenty twenty-five episode.
00:05:03.358 --> 00:05:05.004
It's how an agent gets hands.
00:05:05.728 --> 00:05:14.954
Essentially, an MCP server gives an easy way for an agent to query and work with data on another system.
00:05:15.454 --> 00:05:42.949
So essentially, I had Claude Code build this little wolf system that pulls in this data on a daily basis and then gives it a way for other agents to then get at and work with that data So then when I want to actually work with this data, all I need to do is open up a terminal, load up, load, load up Claude Code, and ask it a question in English, like which pages lost organic clicks s-since September and did affiliate revenue follow?
00:05:43.649 --> 00:05:44.540
Could be a question.
00:05:45.040 --> 00:05:51.189
And then Claude Code itself will take that question, go to the MCP server to Wilf.
00:05:51.290 --> 00:06:19.870
It will interrogate backwards and forwards, work out what data it needs, get all that data back, save it in-- locally for itself, query it, and build an answer, and then come back to me and tell me what it's found From that, I can then cross-question it, ask more questions about the data it's found, even ask it to visualize that data, create an artifact for me or a PDF or something useful from that data that we can then work further with.
00:06:20.505 --> 00:06:37.846
So it's essentially an incredible way of having a conversational interrogation of all the important data without having to trawl through lots of different dashboards, lots of different systems, without having to pull data manually all the time.
00:06:37.899 --> 00:06:45.382
It's just makes it simple But here's the next part, and this is about making it trustworthy, and it's the rule I mentioned at the top.
00:06:45.882 --> 00:06:55.613
Wilf has a fax file the agent must read before doing anything else This is where we give it the context that it needs to properly understand the data.
00:06:55.776 --> 00:07:13.377
An example being back in September twenty twenty-five when the normally, of course, one hundred flag was removed from Google Organic Search, and it made a big difference to the number of search co-console impressions that search Search Console reports.
00:07:13.877 --> 00:07:22.088
That kind of fact I've given to Wilf as I've worked with it because it came back and said,"Oh, there's this big crash in data," and I said-- And give it that explanation.
00:07:22.088 --> 00:07:23.166
It says,"No, that's fine.
00:07:23.569 --> 00:07:30.494
That happened because of this is what Google did." And it goes,"Oh okay." And it saves it to its context, to its memory, essentially.
00:07:30.846 --> 00:07:34.170
So it knows in future that, oh, yeah, the reason that drop happened was nothing to do with the site.
00:07:34.197 --> 00:07:35.745
It was to do with this change at Google.
00:07:36.100 --> 00:07:44.394
So when these things come up, you can backfeed information to it so that it improves its answers further down, and that's the trustworthy, that's the trustworthy piece.
00:07:44.416 --> 00:07:52.262
'Cause this is why you always still need to question the data like you would if you were working on it by yourself without an agent, like the old-fashioned way.
00:07:52.579 --> 00:07:59.045
You still need to question your data, and then those assumptions, those things that you learn from your data, you need to feed back in.
00:07:59.055 --> 00:08:10.396
Because things that you-- That's context you would remember naturally as a human, you feed it back into it And without that kind of context, the agents can do things like confidently explain traffic drops that never happened.
00:08:10.783 --> 00:08:14.713
With them, it reasons like someone who's worked on your site for years.
00:08:15.213 --> 00:08:16.740
You need to tell it what you know.
00:08:17.370 --> 00:08:20.423
It won't remember otherwise So that's working with my data.
00:08:21.160 --> 00:08:25.985
The second use is everybody else's bots So four things here quickly.
00:08:26.485 --> 00:08:30.810
One, Salience ships a command line tool and an MCP server itself.
00:08:30.949 --> 00:08:37.442
So from the terminal, I can ask which AI bots hit this site this week and which ones got blocked.
00:08:37.923 --> 00:08:41.924
The agent, it'll then call Salience, it'll read the log data, and it'll answer.
00:08:42.384 --> 00:08:49.482
And it-- we built Salience for agents on purpose because this is how I wanted to work with the data.
00:08:50.047 --> 00:09:12.123
We obviously have all the dashboards, all those kind of things, and they're important for certain times, but it's so much more powerful nowadays Being able to have your agent talk directly to your services like that because it can just pull data in and out and you can query it and work with it in ways that it's just so much harder to do if you're just relying on a dashboard only Secondly, browser tools.
00:09:12.259 --> 00:09:16.256
Claude Code can drive a real browser through another MCP server.
00:09:16.601 --> 00:09:19.532
So you can say,"Load this page the way Googlebot would.
00:09:20.215 --> 00:09:22.518
Does the main content render without JavaScript?
00:09:22.605 --> 00:09:30.176
Screenshot it, please." That used to be a Chrome extension, a dev tools panel, and lots of squinting and cutting and pasting.
00:09:30.482 --> 00:09:35.301
Basically, you can now say to your agent,"Go to this website, perform this action for me."
00:09:35.801 --> 00:09:38.696
Thirdly, keyword research straight into the brief.
00:09:38.697 --> 00:09:41.966
So
Keywords People Use has its own MCP server.
00:09:42.456 --> 00:09:44.500
I definitely did an episode on that in the past.
00:09:44.534 --> 00:09:47.503
We were one of the very first places to do any kind of MCP server, I think.
00:09:48.277 --> 00:09:55.167
But with that MCP server, the agent can pull the people also ask questions and related searches for a topic without having to leave the terminal.
00:09:55.168 --> 00:09:57.092
You don't actually have to go and use our website anymore.
00:09:57.092 --> 00:09:59.033
You can just use the MCP server.
00:09:59.037 --> 00:10:02.322
So you can say to your Claude,"Go get me related searches.
00:10:02.326 --> 00:10:21.498
Say, get every question people ask about heat pump running costs and cluster them by intent, and draft the FAQ session-- section with one source per answer." And then this will go and research, get that data from Keywords People Use directly, pull it in, work on that first draft, give it to you ready for you to work on.
00:10:21.998 --> 00:10:35.591
And that is just amazingly powerful how quickly it can pull that data in, work with it, and it just saves you so much time and gets you much closer to having the published content that you want And fourth your websites are code.
00:10:35.615 --> 00:10:58.711
Our sites are code, so the agent can work and fix o-sites directly so now you know, every page or any page that we want to be worked on, we can just tell the agent what it is that we want done, what we need changing, where to get the data from All those kind of things And you can then send to it, open up a change for me to review.
00:10:59.475 --> 00:11:02.596
You can check it and then just click Merge when you're happy for it to go through.
00:11:03.049 --> 00:11:22.488
So You can just go end to end from querying the data, finding out what the problem is, getting the data you need for the new content that you need to work on, and getting it onto the site all from within your agent, all from within the terminal, all by working with Claude Code.
00:11:22.988 --> 00:11:30.275
So back in the old days, a technical SEO audit used to end with a document someone else might act on or that you would act on yourself later.
00:11:30.671 --> 00:11:39.666
Now, it just ends with a pull request that Claude's done for you, that you can go and review, work on, and then decide when to merge and publish.
00:11:39.668 --> 00:12:00.873
So yeah, it's just a completely different process, much more streamlined, much more efficient And once you've got the hang of it, the learning curve is actually less than trying to learn all the different tools that you need to work with otherwise to p- to achieve the same result And the third use, that was the second use.
00:12:00.873 --> 00:12:06.399
So the third use is if you heard the link building series you know my view.
00:12:06.905 --> 00:12:12.216
Build something that's worth citing, so tools, data, awards, all that kind of thing.
00:12:13.041 --> 00:12:19.985
And
this summer, a journalist at a national newspaper in the UK, they wanted a data set from one of the, that one of my sites holds.
00:12:20.485 --> 00:12:29.884
And with Claude Code, I got them together a doc- document package, document data package, which covered every postcode, that's a zip code if you're in America.
00:12:29.890 --> 00:12:38.006
In the UK, that's over one and a half million postcodes worth of data in just one session with Claude Code.
00:12:38.256 --> 00:12:53.623
I also had it pull the regulator's public register, that's about another 11,000 records, into something that we could actually query and publish from for them, and pull that all together in a special data package that was branded for this newspaper.
00:12:54.087 --> 00:13:03.205
It had dedicated tools within it that the journalist could use to query that data and get all the stats together they needed for the piece of work they were working on.
00:13:03.715 --> 00:13:11.985
And it was just an hour or so session with the agent gathering the data, creating the tools, creating the data package, and sending it off.
00:13:12.485 --> 00:13:19.690
Two years ago, each of the bits of the package in there, not just the whole package itself, would've been a week of a developer's time.
00:13:19.690 --> 00:13:20.899
So it just wouldn't have happened.
00:13:21.168 --> 00:13:22.416
We couldn't have afforded to do it.
00:13:22.416 --> 00:13:23.341
Wouldn't have had the time.
00:13:23.794 --> 00:13:26.144
Now, it's a session in an afternoon.
00:13:26.287 --> 00:13:27.878
So these things do happen.
00:13:28.378 --> 00:13:33.019
And that's really the, what I'm talking about in the whole of the last three episodes here in one big example.
00:13:33.023 --> 00:13:36.078
The bottleneck isn't the building anymore.
00:13:36.428 --> 00:13:38.381
It's not the time to build that's the bottleneck.
00:13:38.654 --> 00:13:42.804
It's deciding what's worth building that is the bottleneck So that's it.
00:13:42.817 --> 00:13:49.625
That is the four rules that I've learned over the past year or so working with agents like this is.
00:13:49.815 --> 00:13:54.892
So it's one You've gotta watch them because they will fill a gap with something sounding plausible.
00:13:55.317 --> 00:13:56.739
Always make it show it's working.
00:13:56.754 --> 00:14:00.409
Always interrogate the number that it comes back with.
00:14:00.525 --> 00:14:01.857
Ask for the query behind it.
00:14:02.225 --> 00:14:13.914
Keep a second source for it to check against Secondly, write things down for it or speak to it to tell it to write this down for itself, these instruction files, the facts files.
00:14:14.649 --> 00:14:21.849
Every session will start blank unless you or the agent itself has written it down somewhere for it to be shown and to refer to again.
00:14:22.575 --> 00:14:27.716
Thirdly, don't give it the power to send, publish, or deploy on its own.
00:14:28.216 --> 00:14:32.931
You can have it draft, have it propose, you can have it open a change, but always make sure you review this.
00:14:33.873 --> 00:14:37.429
The agents that work on Salient, they can't deploy, only I can deploy.
00:14:37.796 --> 00:14:41.984
My inbox agent can draft a reply, but it's never been allowed to send one.
00:14:42.028 --> 00:14:48.249
I always run over the draft myself, make changes where needed, and press send myself.
00:14:48.749 --> 00:14:51.884
And fourth, start smaller than you think, okay?
00:14:52.381 --> 00:15:00.217
Don't just go and say to it,"Audit my site." Start with one question, one page template, one export, then widen it out.
00:15:00.238 --> 00:15:02.791
With all these things, start small and then widen out.
00:15:03.504 --> 00:15:15.405
If you were to do one thing this week, I would suggest you First of all, download ClawCode, and then secondly, export the last, say, 16 months of your search console performance data to a folder.
00:15:16.139 --> 00:15:27.009
Install ClawCode, open that folder, and ask it one question such as this: Which pages lost clicks after September 2025, and is the drop in impressions or in clicks?
00:15:27.626 --> 00:15:28.464
That kind of thing.
00:15:28.706 --> 00:15:36.361
Then read what it did, read what it says, and then open that conversation with it and see what you can learn from it So that's it for today.
00:15:36.596 --> 00:15:40.549
Hope you found this episode useful, and until next time, take care and I'll see you later