Last week I published a blog post, generated its hero image in four sizes, updated the sitemap, committed only the right files, and pushed it live — in about six minutes, from one chat. No designer, no editor, no ops person. There isn't anyone else. It's me and Claude Code.
I keep seeing "I tried 100 Claude skills" roundups. I didn't try 100. I run a one-person software studio, and I use a small number of skills hard, on work that has to ship and sell. So this is the honest version: the handful that actually earned their keep, what they changed, and — the part most roundups skip — where they stopped.
The setup
I sell a paid Maya animation plugin. I write the blog that markets it. I do the SEO, the research, the marketplace listings, the deploys. Alone. That mix is the point: the value of an AI skill isn't a demo, it's whether it survives contact with a real deadline and a real customer.
Three earned their place. One thing they all share: they're good at the work, and useless at the judgment. That line matters, so I'll keep coming back to it.
1. The writing skill — one voice across 250 posts
The problem it solved is boring and real: consistency. My blog has over 250 articles. Left alone, an AI drifts into the same mush every time — "in today's fast-paced landscape," "game-changer," a tidy summary that says nothing.
The writing skill I use rewrites that instinct. Concrete rules, applied every time: lead with the actual example, not the throat-clearing. Short sentences. No hype words. Never invent a fact or a number. End with something usable, not a soft recap.
The difference is easy to see. Before, an intro to a tools article would open with "Animation software has evolved rapidly in recent years." After, it opens with the animator's actual pain: you set a pose, scrub two frames, nudge tangents so the in-between doesn't pop, and then you do that ten thousand more times. Same topic. One of them a reader keeps reading.
Where it stops: it has no ideas and no facts of its own. It's a voice, not a brain. If I feed it nothing, it writes nothing worth reading. Every real detail — the price, the feature, the tradeoff — I supply and I check. The skill makes my input sound like me. It does not replace my input.
2. The research skill — it caught me being wrong
This is the one that surprised me.
I needed to know where a solo developer can actually sell a Maya plugin in 2026 — which marketplaces are open to new sellers, which accept software and not just 3D models, which quietly shut down. Easy to get this wrong from memory, and I did have wrong assumptions going in.
The research skill fans out: it runs several searches at once from different angles, pulls the primary sources, extracts specific claims, then tries to refute each claim with independent checks before it trusts them. That last step is the whole point. It's not "summarize the internet." It's "assume you're wrong until three checks say otherwise."
It corrected me twice in one run. I assumed one big marketplace still ran an open seller program — it had shut that down months earlier. I assumed a network paid a certain way — the fee page said otherwise. Those are exactly the mistakes that cost you a week if you act on them.
Where it stops — and I want to be honest because the roundups never are: on that run, the verification step hit a usage limit partway through and couldn't finish checking every claim. It handed me the raw findings with the sources and told me plainly that it hadn't verified them. That's the correct behavior, but it means the skill is a research assistant, not an oracle. I still read the sources for the calls that mattered.
3. Claude Code as the ops layer — until it hits your name
The third one isn't a single skill; it's the environment doing the unglamorous work. Processing an image into a JPG and three WebP sizes. Staging exactly ten files and not the private ones. Writing a commit, rebasing cleanly when an automated process pushed at the same time, verifying the page was live afterward. Driving a browser through a seller-onboarding flow — the profile, the forms, the parts that are pure clicking.
For a one-person shop, this is the difference between shipping weekly and shipping when I get around to it.
And here's the boundary, sharp and correct: the moment a step needed me, it stopped. Accepting a legal agreement. Entering bank details for payouts. Choosing a tax status. Signing a form under penalty of perjury. Every one of those, it laid out clearly and then handed back to me. It didn't click "I agree" on my behalf, and I wouldn't want a tool that would. The work is delegable. The liability is not.
What it did not do
This is the section I wish those "best skills" lists included, so here's mine.
It did not move my domain's authority. I built links, submitted to directories, got a listing merged into a popular resource list. My Domain Rating didn't budge. Because most of those links are nofollow, half aren't crawled yet, and authority is a slow, lagging thing no amount of prompting accelerates. A skill can do the outreach. It cannot manufacture the outcome.
It did not replace judgment. It told me a marketplace was technically open, then told me — honestly — that listing my plugin there would mean miscategorizing it as a 3D model, which would probably get rejected. The facts were the skill's. The decision to walk away was mine.
It did not make my product good. It made my shipping fast. Those are different things, and confusing them is how people end up shipping fast garbage.
The honest shortlist
If you run something small and alone, and you want the version without the hype:
- A writing skill with hard rules — for consistency across everything you publish. It amplifies your voice; it doesn't supply one.
- A research skill that verifies before it trusts — for decisions where being wrong is expensive. Read the sources anyway on the calls that matter.
- The coding and ops environment — for the repetitive shipping work, with a clean stop at anything that needs your signature.
The through-line is simple. These skills are extraordinary at the work and deliberately helpless at the judgment. Pointed at real, repetitive tasks, with a human who still reads the sources and makes the calls, they let one person do what used to take a team. Pointed at your judgment, they'll happily let you skip it — which is exactly when you shouldn't.
I don't have a team. I have this. After shipping real work with it, that's the honest trade: it gives me back the hours I'd lose to the mechanical parts, so I can spend them on the parts that are actually mine.
