I Spent 80 Hours Writing 14 Claude Skills. Here’s What That Time Actually Bought Me.

I came to Claude after two years on ChatGPT. The skills started as a curiosity. Six months later they had quietly rewritten how I run two small companies.

I used ChatGPT for almost two years before I switched. The last twelve months of that I was on the $200-a-month tier, running it for almost everything—supplier research, listing copy, draft articles, the whole operational layer of two small companies.

Claude was on my radar but I was not interested. I had a setup that worked, more or less. Then a different kind of YouTube video started showing up in my feed. Not the breathless “ChatGPT vs Claude” comparisons, but quieter ones from people who had clearly been running both for a while and had landed somewhere specific. I watched a few. I bought the $20 plan to try it out. Within three weeks the usage cap was bothering me, so I moved up to the $100 plan, where I have been since. That was the real signal — when I noticed I was changing my plan to fit the tool, not the other way around.

Somewhere in those first weeks I started reading about skills. The idea, once I understood it, was simple. You write the logic of a task in plain English — your standards, your sequence, the specific decisions you want made — and the model follows it. You stop having to re-explain yourself in every new conversation. The skill stays. The work compounds.

I started building. I pulled inspiration from skills people had published on GitHub, took the bits that fit my work, and started writing my own. After the first one came together, I noticed something I had not expected. Once you see how the structure works, you start seeing skills you should write everywhere — every recurring task, every standard, every situation where you wished the model just knew how you wanted it done. The list does not run out. There is no bottom. You finish one, look at it, and immediately think this is good but it would be better if I added one more thing.

Six months later I have written fourteen of them. The total time, conservatively, is around eighty hours of focused work. The first three were the most expensive. By the tenth I had a method. By the fourteenth, I could write a working first draft in two hours, polish it in four. From scratch the cost is now about two days. If I am adapting something useful from GitHub it is closer to one.

This is the part of the AI conversation almost no one is having. The rest of it is all about how much time AI saves you. The math sounds great. The math is also, in my experience, somewhere between exaggerated and false, at least if you are trying to build something more durable than disposable content.

The eighty hours did pay back. They are paying back now, every week. But the bigger thing those hours did is something the popular pitch never mentions — they made me look at my own business in a way I had been avoiding for fifteen years.

My method, since I should be specific

Before going further, here is how I actually write a skill, because it took me a while to land on a process that works.

I start with a roadmap. Not the content — just the shape. What is the task. What are the inputs. What does a good output look like. What are the edge cases. What should the model never do. Roughly half a page of bullet points before I write a single instruction.

Then I run a SWOT on the task itself. Strengths — what is already working when I do this manually. Weaknesses — where I keep making the same mistakes. Opportunities — what would unlock if this ran reliably. Threats — what could go wrong if the model gets it wrong unsupervised. This step is the one I almost skipped on the first few skills, and almost every problem I had later traced back to a SWOT I had not done.

Then I fill the body. The voice rules, the thresholds, the templates, the examples, the negative examples. This is the slow part. This is where the work actually lives.

Then I test. Not with an abstract test case. With a real piece of my own work that I already know the right answer for. If the skill produces output that matches what I would have done — or close to it — the skill is working. If it does not, I read the output carefully and look for which part of the skill is missing the instruction that would have closed the gap. I add that instruction. I run it again.

Most of my skills got their final shape on the third or fourth iteration. The first version is always thin. The second version reveals what I forgot to put in. The third version is when I start trusting it.

After that, I keep them on a maintenance schedule. The operational skills get reviewed every three or four weeks, because the marketplaces and the suppliers and the search algorithms keep moving and the skill has to move with them. The strategic skills I update less often, maybe once a quarter, because the underlying logic is more stable.

I am also being honest with myself about something. There is no finished version of any of these. Every time I review a skill, I find at least one improvement I could make. The phrase that keeps coming back is this is good, but it would be better if — and that sentence never seems to run out.

The instruction set is the asset. The tool that runs it is almost a side effect.

Five of the skills, and what they actually do

The abstract argument I am about to make would be empty without specifics, so let me describe five of the skills I run every week. These are the ones doing the most work for the business right now.

The supplier-vetting skill. This was one of the first I wrote, because the bottleneck was obvious. I deal with Chinese suppliers constantly — 1688, Taobao, Tmall — and the cost of a wrong choice is high. A bad supplier will not just sell you a counterfeit product. They will damage your eBay account, your feedback score, and your buyer trust in ways that take months to repair. I needed a faster way to get to the suppliers actually worth talking to. The skill takes a screenshot of any supplier page and runs it against a fixed three-criterion filter — repeat-purchase rate, store age, feedback rate, with an automatic red flag if the price looks too low for a young store. What used to take me an afternoon of comparison work now takes thirty seconds. More importantly, it has stopped me from placing several orders I would have placed under time pressure without it.

The ebay-listing-generator skill. This one is the most measurable. Before the skill, I was writing eBay listings the way I had been writing them since 2008 — by hand, with a half-remembered template, with item specifics filled in inconsistently across the catalog. The skill encodes the structure I want every listing to follow, the title formula, the item-specifics conventions, the condition language, the photo-position rules. A listing that took me an hour now takes ten minutes. The harder number to share, but the more important one — listing volume and consistency since the skill went live have lifted sales close to 88% above where they were before. I will not pretend the skill is the only reason. But the consistency it produces is doing something the manual approach was not.

The price-research skill. This one I use earlier in the cycle, when I am deciding what to source in the first place. It pulls real-time UK market prices from eBay sold listings and competitor stores, looks at the spread, and tells me whether a product fits the kind of margin I need. I use it for two opposite cases. Sometimes I am chasing a hot product, where margins are tight but volume is high. More often, and where it has paid off most, I am hunting niche products — items that may sell in low volumes but carry healthy margins because almost no one else is bothering. The skill flags the second category in a way I would have missed scrolling through eBay manually. Several of the products on my shelf right now came from a price-research output I would have walked past.

The competitor-intelligence skill. Different problem, similar pattern. I want to know what specific competitors on eBay UK are doing — what they are listing, what they are quietly dropping, where their stock is moving, where it is sitting. The skill takes a competitor’s storefront and produces a structured read of their current behaviour: which categories they are leaning into, which they are pulling back from, which products are showing low stock in a way that suggests intent rather than accident, where the gaps in their range are. The output is not magical. It is what I would produce if I sat with their storefront for two hours and took careful notes. The skill does it in five minutes, and it does it consistently across competitors, which is the part I could not maintain manually.

The b2b-outreach skill. The slowest-cooking of the five and probably the most important long-term. My business has two sides — the consumer side selling on eBay UK and OnBuy, and the wholesale side supplying UK resellers from Hong Kong stock. The wholesale side runs on B2B relationships, and B2B relationships run on consistent, well-pitched outreach. I was bad at this for years, in the specific sense that I would write a great message to one prospect, then write a worse one to the next, then write something half-formed to the third because I was tired. The skill encodes the pitch logic, the segmentation, the message structures for LinkedIn DM versus cold email versus WhatsApp versus WeChat — different languages where appropriate. The result is that my outreach is now uniformly the version of itself I would have written on my best day, every day, in every language the relationship requires. The network has grown. The supplier conversations have grown deeper. I expect this skill to compound for years.

These are five out of fourteen. The other nine are smaller — content production, contract review, dead stock diagnosis, negotiation prep, morning market briefings — but they share the same shape. Each one is a working protocol that captures a specific operational decision I had been making by feel for years. The thing they have in common is that none of them feels like AI doing my work. They all feel like me doing my work, faster.

The thing I did not realise about my own work

There is a strange thing that happens when you have been doing the same work for a long time. You stop seeing it. Not in the obvious sense — you still do the work, you still get the results — but in the deeper sense that you stop being able to describe what you are actually doing. The decisions get faster, the patterns get cleaner, and the explanation of why you do things the way you do quietly disappears, because you no longer need to explain it to yourself.

For fifteen years, this was my situation. I made hundreds of decisions a week — about suppliers, about pricing, about which listings to push and which to leave alone — and almost none of them passed through conscious words. They felt like instinct. They were not instinct. They were patterns I had built up so quietly that I no longer remembered building them.

Writing the first skill forced me to put it into words. It was uncomfortable in a way I had not expected. Every time I sat down to describe how I check a supplier, or how I write a product description, or how I decide whether a piece of writing is finished, I would write three sentences and then stop. I could not fill in the next sentence. The choices I had been making for years turned out to be based on patterns I had never named out loud.

The supplier-vetting skill, again, is the cleanest example. The first version I wrote told the model to “check feedback rate, store age, and other quality indicators.” That was not a skill. That was hand-waving. When I went back to fix it, I had to ask myself what other quality indicators actually meant when I was looking at a 1688 page. The answer turned out to be specific numbers I had been checking against without thinking — repeat-purchase rate above 40%, store age above three years, feedback rate above 80%, automatic red flag if the price was much lower than market on a young store. I had been applying this exact filter for fifteen years and never written it down. Most of my real decisions were happening in the gap between looks fine and something feels off, and the only way to capture that gap was to make myself name the numbers I had been checking unconsciously.

The same thing happened with the listing skill, the price-research skill, the competitor-intelligence skill, the b2b-outreach skill. Each one started as a vague description and became, over multiple drafts, a real working protocol with clear thresholds, clear exceptions, clear edge cases. Each one taught me something about my own work I had been doing on autopilot.

This is the experience most people running AI workflows seem to be skipping. They use a one-size-fits-all prompt, get generic output, and decide either that AI is amazing or AI is overrated, depending on whether the generic output happened to fit. What they are missing is that the most valuable thing the tool can do is force the kind of self-examination that produces a usable instruction set in the first place. The instruction set is the asset. The tool that runs it is almost a side effect.

The pushback

Eighty hours is a lot of upfront work. The reasonable response is: skip the skill machinery, write better prompts as you go, use Claude or ChatGPT in conversation, save the eighty hours and start producing right away.

I tried this approach for the first three months. It worked, in a limited way. The problem was that every conversation started over. The model did not remember what I had told it last time about my voice, my standards, my edge cases. So I would re-explain the same context every time I started a new task — which is its own kind of repetitive labour. Worse, my explanations drifted from session to session. I would emphasise different things on different days, and the output reflected those drifts. My work became less consistent over time, not more.

The skills fixed this. Once written, the standard holds. Every article goes through the same instruction set. Every listing follows the same template. The consistency itself is valuable, separately from any time savings, because the work compounds in the eyes of buyers and readers — they recognise a pattern, they trust it, the pattern becomes part of how they identify what I do. You cannot get this from in-the-moment prompting. You only get it from doing the upfront work.

What this changes

The leverage point, after all of this, is not the tool. It is what you put into the tool. Generic prompts produce generic output. Twenty years of carefully written-down judgment produces work that carries twenty years of judgment — at speeds that would be impossible without the tool, but at quality that requires the years underneath.

This is the part most people will not do, because it is uncomfortable and slow and produces no immediate reward. The shortcut version of AI will keep being popular. The deeper version will stay rare. The gap between them, in what they produce over years, will keep widening.

A question, because I am genuinely curious how other people experience this. If you have spent serious time building reusable workflows in any AI tool — Custom GPTs, Claude Skills, n8n workflows, anything — did the process of building them teach you something about your own work that you had not put into words before? My guess, based on how this has gone for me, is that the act of writing the workflow is more valuable than the workflow itself. But I do not yet know if that is true everywhere or just true for the kind of work I do.

I have spent twenty years on the supply side of consumer electronics, and the last six months building a Claude-based system for running two small companies. Over the next few weeks I will be releasing one of these skills as a free download, along with a short walkthrough of how I built and tested it. If you want to be on the list when it goes out, follow me here — that is the only place I will announce it.