This is Part 2 of a three-part series on running Claude as my office assistant. Part 1 covered where it’s actually paying off. This part covers what sits underneath that: the backups, the second brain, and the skills I’ve made mine. Part 3 is the honest part: the pitfalls, the questions I get asked, and where I take this next.
Part 1 covered the wins: the content pipeline, the CRM, the daily briefing, the SEO work, the multi-site RAMS build. This part is about what's underneath all of it, the parts that aren't glamorous to write about but are doing more of the actual work than anything in Part 1.
None of what's in this part would have happened in an ordinary chat window. Asking a chatbot questions back and forth is where most people's experiments with AI start and stop, and it's useful for that, drafting an email, checking a figure, sense-checking an idea. But it's not where the actual leverage is for a business. That came from moving to Cowork instead, letting it open a terminal, write and redeploy a script, work inside a shared Google Sheet, manage files on my NAS, keep a task running on a schedule without me sitting there prompting it step by step.
That's the realisation two months in: the value isn't a clever question getting a clever answer, it's handing over enough actual autonomy that it can go and do the thing, then me checking the result. Everything in the rest of this part, the backups, the delivery system, the second brain, the skills, came out of that shift, not out of a longer or cleverer prompt.
Two jobs that turned into one, and I'll keep the technical bit plain because that's actually how Claude talked me through it. First was ensuring that my back up system was robust. I asked Claude to help lock down my backup setup and looked at pulling the holes that I hadn't seen. The stand alone hard drives, Synology NAS and the extra storage drive attached to it, step by step, in an order that made sense: turn on an extra login check so a password alone can't get anyone in, sort out my home internet so it has one fixed IP address instead of a shifting one, only open the single door in that the NAS actually needs, and swap the "this site isn't secure" browser warning for a proper certificate. I didn't need to understand networking, I just to followed the steps in order and asked when something didn't make sense.
Once that groundwork was in, I used the same setup to build a client image delivery system that doesn't touch WeTransfer at all. I wanted a 'in-house' delivery mechanism that fit what I was going to do anyway and just open the resultant project files up to the client via a password on my website. Final high-res files now get served from the NAS through a branded, password-protected page, thumbnails so the client can see what they're getting, a proper download-all button, rather than a link that expires in a week and a nagging email I have to watch for. It's got its own secure web address, locked down the same way, and the password switches itself off once the job's done. This keeps everything branded and under the gregharding.co.uk banner. Some company IT firewalls won't allow access to WeTransfer or similar, so this should get around this limitation.
The Drive connection holds what I call the business's second brain: a wiki of standard operating procedures, rate decisions, client and tool notes and open questions, kept current so any Claude session already knows how the business runs instead of starting from nothing. It took a solid two to three days of focused work to build it out to the point where I trusted it. Ask which single bit of this whole setup has paid off the most and I won't hesitate: it's this, ahead of the content workflow, the CRM, the SEO work, all of it.
I was properly excited setting this one up, yes, I'm a geek about this stuff. It asked a myriad of questions about how the business actually operates, and I couldn't hide behind a vague answer if I actually wanted it to help. For those two to three days I was very much working on the business rather than in it, something I'd tell anyone running their own show to make time for, because it doesn't happen by accident. It would have been lovely to have this when I started out, but the business has grown and morphed and niched down since then, and the procedures and fixes for all of that just lived in my head, nowhere else. That changed all at once, and it was a real relief to watch it actually working in practice rather than staying a good intention.
What surprises me most is how it feels, not just what it does. Having it there makes running this business feel less like doing it entirely alone, there's something to bounce an idea off before committing to it, rather than deciding alone at 11pm and hoping I've thought it through. On the technical side, it's like having an IT person built into the business, sorting DNS, certificates, port forwarding and broken scripts without me having to learn it all from scratch.
Skills are packaged instructions telling Claude how to approach a task, and out of the box they're generic. The value came from personalising them: a brand voice skill built from a full review of my site and YouTube channel (first person, never "we," no bare superlative without a real proof point, none of the stock phrasing that makes AI-drafted copy sound AI-drafted), and a matching SEO skill pinning down which data sources I actually trust. More recently I've started building small, task-specific agents on top of those skills, ones that just run a single job like researching an enquiry, rather than me explaining my business from scratch every time.
Some of the fixing wasn't flattering. A case-study-writing skill I picked up had started life built for a software agency, its first step telling Claude to "explore the client's codebase," which makes no sense for a photography business and would have quietly produced nonsense if I hadn't caught it. I rewrote that step into a proper intake process pointed at my brand voice skill. Separately, a skill covering enquiry checks referenced a scheduled task that, on closer look, didn't actually exist yet, describing a process as already running when it wasn't. I had to build the real thing before the skill's own description was true.
Finding new skills to try wasn't the hard part. Ideas came from all over: seeing how other small businesses and photographers were actually using Claude on Instagram, browsing ready-built skills shared on GitHub, prebuilt integrations for WordPress. Installing one takes minutes, no different to installing a plugin. That ease is exactly the risk. A skill that isn't built for how you actually work doesn't usually fail loudly, it just quietly produces output that looks plausible and isn't.
The clearest example was a finance skill I picked up expecting help with basic invoicing and bookkeeping questions. It turned out to be built for US companies, US tax categories, US invoice conventions, none of which hold for a UK sole trader, and nothing about installing it flagged that up front. I only caught it because I asked Claude to actually review everything I'd installed rather than trust my own read of what looked useful on the shelf, and that review is what led me to disable and uninstall a handful of skills outright, the finance one included. It's the same rule as everywhere else in this series, poor data in, bad data out, except here the poor input wasn't something I typed, it was baked into the skill itself before I'd given it a single fact about my business.
The same restraint applies to integrations, not just skills. Not everything with an MCP connection available needs actually connecting. A tool being able to plug in isn't a reason to plug it in, and every connector I add is one more thing that needs the same scrutiny as a skill does, not a free upgrade just because it's on offer.
Part 3 is the honest one: what's actually gone wrong (a spreadsheet corruption bug, a proposed CRM rebuild that went nowhere, and the ongoing worry that my CRM is only as good as the information I give it), a straight FAQ answering the questions people keep asking me about this, and where I take it from here.