page-foundry

One brief in. A page that argues its case out.

npx skills add taylorbanks/page-foundry

The homepage above was built by page-foundry. That is the demo.  ·  Live site

page-foundry is a Claude Code skill. It turns one brief into a finished homepage, landing page, or sales page: written in your voice, fast, accessible, honest about its proof, and built so it never reads or looks AI-made. Keep the static HTML, or take the design package to any tool you like.

You already know what an AI-built page looks like

You can spot one in about two seconds. So can everyone else, which is the problem. There is a whole shared vocabulary for it now: the gradient, the same three feature cards, the typeface everybody uses.

Here is what is strange about that vocabulary. Go read any of the catalogues of AI page tells. Every item on them is a font, a color, or a layout. Not one is about what the page says.

So the pages get fixed, and they still do not work. The look improves and the signups do not, and now the failure is invisible, because the thing everyone knows how to name has been handled.

The part nobody names

Ask a founder with a page and no signups what is wrong and you get some version of "I don't know what's turning people off." Ask their peers to look, and the answers are never about design. They ask where the pricing went. They say they cannot work out what it actually does, or who it is for, since "everyone" is not an audience. They want to know what makes it different from the dozen others they could pick instead, and where the proof is that any of it works.

Those are the same failure the visitor describes from the other side, in the words they use when they leave: I can't tell what this does.

The reason it is hard to catch in your own page is ordinary and well understood. The page makes sense to you because you already know what the thing is. You cannot un-know it to read your own copy cold. That is a property of expertise rather than a personal failing, and it is why the fix has to come from outside your own head.

An agent will not supply it either. Asked for a landing page, it has to decide who the page is for, what it competes with, which objection to answer first, and which proof leads. Nothing told it any of that, so it fills the gap with the average of every page it was trained on. That average is where the sameness comes from, and it is why two runs of the same prompt drift: the average moves when the wording does. One missing input, both symptoms.

What page-foundry does before anything gets designed

It writes the spec first. Five questions, five files.

Then the design direction, the copy, the build. The design phase runs on real craft references and a deterministic detector (impeccable) that persists a design system for the property and then rejects anything off it, which is what keeps the look from drifting back to the gradient-and-three-cards default. Each stage opens the file the stage before it wrote, so the headline traces back to a claim, the claim traces back to a quote, and the quote traces back to a link you can open.

That chain is the product. When it holds, the page argues a case instead of describing a product, and the argument is made of things a real buyer actually said.

"So it is a prompt that says do the positioning first"

Fair question, and it is the right one to ask, because that describes most of this category. The difference is what happens to work that does not hold up.

Things that stop a page here:

• Copy that trips the voice scanner, which reads its rules from a file you edit, so page forty sounds like page one
• A page whose rendered text has drifted from the approved copy, checked by diff rather than by trust
• A testimonial, a metric, a command, or a screenshot of something that did not happen
• A claim left standing with no proof beside it, which becomes a marked gap instead of a nice sentence
• A design that trips the visual anti-pattern detector, run over the built page against the property's own design system
• Contrast, keyboard access, or markup that fails WCAG 2.2 AA
• A hero that needs 400KB to say hello

There is a conversion audit too, and it works in a way worth describing: the score of record does not come from the agent that built the page. It comes from a fresh one handed exactly two things, the page as a visitor meets it and the brief. Never the spec, never the reasoning. A page that only makes sense to its author scores badly, which is the entire point, because that is the failure this whole document is about.

It is slower. It will refuse to hand you the page a one-shot prompt would have given you five minutes ago. If your proof is thin it says so and builds around what is real, rather than inventing the testimonial that would have looked better.

What you get

Whatever page you need

Sixteen archetypes, each a conversion contract rather than a fixed template: open source project, SaaS homepage, campaign landing, pricing, comparison, docs, waitlist, event, agency, e-commerce, mobile app, course sales, membership, newsletter, personal site, and a launch changelog. Section order follows how your buyer raises objections instead of a numbered slot, and a page that straddles two archetypes gets a merged contract.

Three ways to run it:

• build: one brief in, a finished page out.
• explore: contrasting design directions first, you pick, then it builds the winner.
• handoff: a complete copy-and-design package for Claude Design, Open Design, Codex, Gemini, or whatever you build with.

Install

# recommended: the skills CLI (installs for whatever agent you run)
npx skills add taylorbanks/page-foundry

This is the way to install an agent skill: it works across Claude Code, Codex, Cursor, and the rest, installs to the right place for your agent, updates with npx skills update, and is how people find the skill on skills.sh.

Other channels:

• Claude Code plugin: claude plugin marketplace add taylorbanks/page-foundry then claude plugin install page-foundry@page-foundry
• claude.ai: upload the .skill file from the latest release
• One branded command (Claude Code / agents dirs): npx page-foundry install — a self-contained installer if you would rather not use the skills CLI. It does not register on skills.sh, so prefer npx skills add unless you specifically want this.

--project        install to ./.claude/skills instead of ~/.claude/skills
--agents         install to ~/.agents/skills
--dir <path>     install under <path>/page-foundry
--with-commands  also write a /page-foundry command stub
--dry-run        print actions without writing
--force          overwrite an existing install

The installer, bin/page-foundry.js, is one dependency-free Node file with no network calls and no telemetry, built on Node standard modules only. It preserves a customized voice.md across updates. Read it before you run it.

First run

Run /page-foundry with no arguments and it prints what it can do, then stops and asks. Nothing starts until you say so.

When you are ready, say "set up my voice". A short wizard writes your voice rules, and those same rules drive the scanner, so your writing guidance and the enforcement cannot drift apart. Until then a neutral default applies.

Read it before you trust it

page-foundry is new, and few people have installed it. There is no download count here worth quoting, and inventing social proof for a tool whose main promise is refusing to invent proof would be a poor start.

What is offered instead is that you can check it yourself in about a minute:

• It ships two executables, scripts/voice_scan.py (the voice gate) and scripts/run_audit.py (the orchestration gate), both standard-library Python with no network access, no subprocess calls, and no dependencies. They are short. Read them.
• The npm installer is the same story: zero dependencies, Node built-ins, no network, no telemetry.
• Every skill it leans on is named and linked below, so the capability claim is checkable against projects that do have adoption.
• A run leaves its work as files on disk. The brief, the buyer quotes with their sources, the message hierarchy, the spec. You can read what it decided and why, after the fact, rather than taking the process on faith.

Companion skills install only from the pinned sources in the skill's own table, only when you approve, never from search results. See SECURITY.md for reporting.

One more piece of honesty about fit: the research behind this positioning clears its evidence bar for one buyer, the solo builder shipping their own product. The archetypes cover more ground than that, and if you are an agency or a maintainer, parts of this are still inference.

Built on

These projects do the heavy lifting. page-foundry does the sequencing and the checking, and it is a lesser tool without any of them. Eight are core, meaning the run does not start without them, because they are what makes the output a page-foundry page rather than a design-tool guess: product-marketing, customer-research, marketing-psychology, cro, copywriting, frontend-design, humanizer, and impeccable. The rest are enhancers that degrade to a condensed fallback when absent.

• marketingskills by Corey Haines: product-marketing, copywriting, CRO, customer-research, pricing, and psychology. It supplies the marketing half of the pipeline (six of the eight core skills).
• impeccable by Paul Bakaus: the design engine. Craft references, a persistent per-property design system, and the deterministic detector behind the render gate. Core: a page whose look was never run through its detector is not a page-foundry page. Its "a clean scan is a floor, never a verdict" principle governs every mechanical check in the skill.
• Anthropic's skills: frontend-design, web-artifacts-builder, and skill-creator.
• web-design-guidelines by Vercel Labs: accessibility, typography, and UX rules.
• humanizer by blader: the "Signs of AI writing" pattern set behind the copy pattern pass.
• gstack by Garry Tan: design consultation, the variant shotgun, and visual review.
• Remotion: hero demo clips and motion, when a page earns them.
• The MECLABS Institute Conversion Sequence heuristic.
• The skills CLI and skills.sh by Vercel.

The eight core skills are required: the run stops at preflight until they are installed (you can override that in chat, but the run is then marked partial and says so). The enhancers are optional; when one is missing the skill falls back to a condensed built-in version of its rules and tells you the run was partial, rather than pretending. Install all of them anyway; the fallbacks are a floor, and the companions are the standard.

License

MIT. See LICENSE.