ReWire skill pack · for Claude Code

Your AI-built site looks AI-built.

Everyone can prompt. The gap between a page Claude generated and a page that looks considered is a few dozen specific decisions — which easing curve, how a pinned section is actually structured, where the 3D sits so it doesn't wreck load time. This pack is those decisions, written down and installed as skills.

Built with it, in a day each. Open any of them and try to break it.

The argument

Same tool. Same day. The difference is what it was told.

Point Claude Code at a brief with no craft constraints and you get the house style of the whole internet: a centred hero, a purple-to-blue gradient, three feature cards, and motion that eases when it should be linear. It is competent and it is instantly recognisable as generated.

Nothing about that is a limitation of the model. It is a missing specification. Given the actual numbers — the type roles, the one accent, the easing vocabulary, the scroll-stage architecture — the same tool in the same afternoon produces something you would put in front of a client.

That specification is the product. Not the videos, not the prompts. The numbers.

What it is

A toolkit, not a course.

A video course about AI tooling starts depreciating the day it is recorded. This is a versioned zip of skills that Claude Code loads and applies while it builds.

  • It runs, you do not watch it

    The skills install into your workspace. `rewire <url>` runs an eight-phase pipeline and stops at five gates for your input. There is nothing to sit through.

  • Updates ship, not re-films

    A versioned zip, not a video course. Updates ship through Gumroad and past buyers re-download.

  • Every rule carries its evidence

    The craft layer names the build each number came from, and says plainly where the builds disagree with each other or fall short of the rule.

What's in it

Eight skills, a working starter, and the craft layer.

  • The taste layer

    The craft skill — the actual numbers behind the sample builds. Type roles, the easing vocabulary, scroll-stage architecture, adaptive 3D quality tiering and the mobile WebGL policy, each with the reasoning and the evidence attached.

  • hydra — adversarial review

    Fans out independent reviewers in fresh contexts, then merges what they find. Runs automatically before the flagship gate. Works on any Claude Code project, not just websites.

  • The eight-phase pipeline

    harvest · analyse · angles · brief · build, plus the rewire orchestrator that runs them and stops at five gates for your input. Every phase is usable on its own.

  • A working starter

    A scroll stage you can serve and scroll immediately: CSS-pinned rather than GSAP-pinned, scrub-driven, with frame-time tiering and a static fallback. The shortest path from the pack to something that looks like the samples.

  • Vendored runtimes

    GSAP, ScrollTrigger, Three.js and Lenis as minified files, with licences. No build step and no framework — this is genuinely all the sample builds use.

  • Spec template and prompt library

    The build spec with its acceptance checklist, and a copy-paste prompt library grouped by what you are trying to fix.

The craft layer

This is the level of detail. On the page, before you buy.

Four rules lifted straight out of the taste skill. Every one of them is checkable against the demo builds — open the source and look.

A scrubbed tween carries ease: 'none'

The scroll is already the timing function. An ease on a scrubbed tween accelerates away from the user's hand and then waits for it, which reads as lag. This is the single most common defect in unprompted AI-written GSAP — it will reach for power2.out every time unless told not to.

Never GSAP pin. Sticky stage plus a spacer.

pin injects a wrapper, takes over the section's layout, and is a reliable source of layout shift and mobile breakage. Pin in CSS with position:sticky over a tall spacer, and use ScrollTrigger only as a progress source. The spacer height is then your timing control.

Scrub is a number, chosen from the moment

Not true, which feels glued. A staged sequence wants 0.45–0.55 so it answers your hand; a parallax band wants 0.6–1.0, because the longer lag is what reads as distance. A hero exit is tightest of all — the user is trying to leave.

Measure frame times, take the median, only step down

The mean lets one garbage-collection spike permanently demote a fast machine. Start at medium until there are ten samples — never start high and drop, because the drop is visible. Quality is a table of settings, not a boolean.

The two modules behind that last rule ship in the pack with their unit tests. Twelve tests, no browser, no canvas, no install step — run them yourself in one command.

hydra

A reviewer that has not read your code yet.

The reason a long build drifts is that the thing reviewing it is the thing that wrote it, carrying every assumption it already made. hydra fans out independent reviewers in fresh contexts, has them work adversarially, verifies the findings and merges what survives. It runs before the flagship gate. It is not a web-design tool — point it at any Claude Code project.

Before you buy

What you need, and who this is wrong for.

Stated up front on purpose. A refund from someone who could never have run it is worse for both of us than a sale that never happened.

You need

  • A Claude Code subscription
  • A terminal you're comfortable in
  • Enough HTML and CSS to read an error message

You do not need to know GSAP or Three.js. That is what the pack is for.

  • You want a website, not a workflow

    If you need a site built rather than the means to build one, the done-for-you service is the better buy — and cheaper than learning this to do it once.

  • You have never opened a terminal

    This is not a no-code product and it does not pretend to be. You'll be running commands and reading errors on day one.

  • You want it to work without you

    The pack raises the ceiling on what you can build. It does not remove the judgement — deciding what looks right is still your job.

Price

One payment. Every future update.

$129AUD

one-off · includes every future update

  • The full pack — eight skills, references, starter, vendored runtimes
  • The taste craft layer, with its evidence and its known gaps
  • hydra, usable on any Claude Code project
  • Every future version, re-downloadable
  • Get the pack

If it does not install, I will fix it or refund you.

Run the install, and if it does not work in your workspace, reply to your receipt and tell me what happened — I will either fix it or refund you. What I will not do is promise the pack will make you money, because I have no way to know that and neither does anyone else selling you something similar.

The zip is markdown and JavaScript. Nothing stops you sharing it, and I have priced it knowing that. What you cannot copy from a leaked folder is the updates, or being able to ask me why a rule is the way it is.

Questions

The things worth asking.

Do I need to know GSAP, Three.js or WebGL?
No. You need a terminal, a Claude Code subscription, and enough HTML and CSS to read an error message. The pack supplies the technique; you supply the judgement about whether the result looks right.
Is this a course? Are there videos?
It is a toolkit. The skills do the work inside Claude Code. Where videos exist they are documentation for the tooling, not the product — the product is the zip, and it is versioned.
What actually makes the sites look different?
A closed set of type roles with a monospace label voice; one accent hue and nothing else saturated; one easing curve declared as a token; scroll stages pinned in CSS rather than by GSAP; and real content instead of placeholder. The craft layer is the specifics of each.
Does it work on mobile?
Honestly: partly, and the pack says so in writing. iOS composites a fixed WebGL canvas with a visible blink, which is a compositing bug rather than a performance one, so a frame-rate probe cannot see it. The starter refuses to create a context on a coarse pointer; the flagship build ships one and recovers from context loss. Neither fully solves it, and the craft layer names that as an open gap rather than hiding it.
Will this get me clients?
I have no idea, and anyone who tells you otherwise is guessing. What it does is raise the ceiling on what you can build and how fast. What you do with that is not something a zip file can promise.
Can I use it on client work?
Yes. Unlimited projects, commercial included, charge whatever you like, no attribution required. You just cannot redistribute the pack itself.
What if I would rather it was just built for me?
That is the other half of what we do, and for a one-off site it is usually the better buy. Start with a free review instead.

Or don’t build it yourself

Want the site, not the workflow?

If you need one site rather than the means to build many, buying the skill pack is the expensive way round. We build these for businesses — start with a free review.