How I built Nexin, a library of Claude skills

The problem it solves

Publishing every day without a content team means writing the same kinds of things over and over: hooks, captions, carousels, threads, pins, posts, sales pages. A chat model will write any of them. It will not remember your brand between sessions, it will happily invent a testimonial, and it sounds like a chat model.

Nexin is built from patterns I use in my own projects, rewritten from scratch as a product: a content department as files you install, so one idea can become a week of posts across channels in your voice, with proof you can actually show.

The decision that holds it together

Every skill starts with the same step: look for the brand files.

The free Core module runs an interview once and writes five files: the brand, the persona (who speaks), the proof ledger, the voice of the customer and the winners. The Product module adds a sixth, the funnel. Every other skill reads those files before it writes a word.

That is what makes a library behave like one system instead of a hundred prompts. A carousel skill and an email skill agree on the offer because they read the same file. A proof rule lives in one place: only proof marked showable is stated as fact, and everything else becomes a bracketed placeholder for the owner to fill. This post was planned with two of these skills, and that rule is why it has no invented numbers in it.

Building the store took a day

The stack is Rails 8.1, PostgreSQL and Solid Queue inside Puma, the same as my other Rails apps, deployed as one container on Railway. Skills, prompts and blog posts are Markdown files in the repo. The database holds only users, orders, entitlements and events.

Two early calls saved time:

• React and Inertia were gone eight minutes after the scaffold. The decision in the plan reads: everything is ERB, simpler and fastest. For a store with a library behind a login, server-rendered pages were enough.
• Delivery stayed inside the app. The spec planned payment links, a separate delivery service and presigned storage links. What shipped is one Rails app with Stripe's embedded checkout and a zip built from what you own, cached after the first build.

The zip has one folder per destination: one zip per skill for Claude.ai, which takes one skill per upload; folders to copy into Claude Code; and the prompts as Markdown for anything else. Prompt bodies you do not own never reach the browser, only their titles.

The Brazil problem

My Stripe account is Brazilian and settles only in reais, so Stripe's automatic currency conversion could not localise US dollar prices, and Brazilian cards could not be charged in dollars. The fix was a currency picker with six currencies that converts at the day's rate. It is the kind of problem no tutorial mentions because most tutorials assume a US account. Choosing a payment provider covers the general version.

Making the skills good took the rest

A library rebuilt on day one

The first library was thirty test skills. Before the end of the first day I replaced them with a library built from a blueprint, then cut and reshaped it to what ships now: ten modules, 101 skills and 221 prompts.

The same day I changed who it was for. The original spec centred on faceless profiles. That became one mode among several: a real person, a persona, a brand or a product can be the speaker, and the persona file gained fields for who speaks, whether they appear on camera and what stories they are allowed to tell.

An eval with a blind judge

I did not want to sell skills on the feeling that they were good. So I built an eval harness: the same brief run headless through Nexin's skills, through Nexin's standalone prompts, and through a baseline of skills from an older internal system of mine, then scored by a blind judge from the same model family.

Round one was humbling:

• The skills leaked internal notes into the deliverable: file names, reasoning, checklists.
• They over-delivered, writing more than the brief asked for.
• Hard bans overrode the brand's own voice. A rule against emoji or a certain sentence shape beat a brand file that used them on purpose.

The fix was a block of delivery rules in every skill: exactly what was asked, a clean deliverable, short notes, and the speaker's voice beats the skill's defaults. Over six rounds the skills went from 6.63 to 7.75 out of 10. The baseline scored 8.08, and the standalone prompts went from 3.82 to 5.32. The eval notes list where the skills still trail: the persona voice, hooks for X, and pins. I would rather publish that than a round number.

289 Threads posts

The Threads skills wrote posts of 150 to 480 characters on one skeleton, with twenty-word replies that read like a consultant. So I studied 289 posts from 21 accounts that do well on Threads. Most of the winners were one sentence under 120 characters. The module was rebuilt around fifteen post shapes and replies mostly under fifteen words, and the judge's score for how human the sales chain read went from 3 to between 6 and 7.5. If you write for Threads by hand, how the Threads algorithm works is the human version of that study.

A fictional teacher

The last test was a full run on a brand that does not exist: a 63-year-old PE teacher I called Carol Whitley. The Core interview ran for 37 turns against a simulated owner whose private notes Nexin never saw, then the whole pipeline ran with nothing edited by hand.

Nexin's own publish gate caught the worst of round one: a sales page selling eight workouts when two had been written, an order bump for a tracker with no file behind it, and a video hook promising a health outcome. The fix went into the funnel file, which now records the offer's terms and what has actually been built. In round two the gate blocked nothing.

Who it is for, and who it is not for

For: creators, personal brands, personas, small businesses and products that publish most days without a team, and want one idea to become posts for several channels in a consistent voice. People with no brand yet are included: that is what the free interview is for.

Not for:

• Autopilot posting. Skills draft and check. They never publish.
• People who will not fill in the interview. The skills are only as specific as the brand files they read.
• Claude.ai on a free plan, if you want the skills there: uploading skills needs a paid plan with code execution. The prompts work anywhere.

The day after I started Nexin, I forked it into Copix, which takes the copy module and serves it as tools any AI client can call. And if you want the method without a product, batching content and building a SaaS with AI are the free versions of how I work.

---

Nexin, if you publish every day without a team

A note on fit first. If you post once a week and enjoy writing every word, you do not need a library of skills. Nexin does not choose your offer for you, and it will not post anything.

Nexin is a content department you install in Claude or ChatGPT: one interview to set the brand, then skills and prompts for Instagram, Threads, X, YouTube, Pinterest, your blog, your offer and your copy, all reading the same brand files.

See Nexin