ai-startup-skills
创业公司做 AI 产品的 7 步 Agent Skill 合集:验证 → BP → 政策 → 路演 → 获客 → 定价 → 注册。
Translation pending
Type a thing. See it pulled apart and numbered, how a factory would make it, and what that would cost — in your language.
OtherGitHub @jamalxcodeOnlineOwner unconfirmed
Have you tried it? One tap is enough. No sign-in needed.
Repository information has not been collected yet.
No updates detected yet. GitHub releases will appear here.
Tell Me Why: type a thing, see how it works. Animated diagrams, exploded parts, real photos and 3D, in any language.
Translation pending
Taken from the repository README
Loading the full README.
Read the full README ↗Live: how.sala.company · part of Sala Company
Type a thing — a combustion engine, scissors, a kettle, how a house is built — and get three answers: how it works (one big exploded view with numbered parts, a few short steps, the real thing in photos and in 3D, in the spirit of the old "Tell Me Why" picture books), how it is made (the factory, in three versions: hand-made, mixed and automated) and what it costs (to set up and to run, and how long each version takes to pay back). In your language.
It's a static site on GitHub Pages. Visitors sign in with their own OpenRouter account, which pays for both the words (Claude) and the picture (OpenAI's image model, also served through OpenRouter). The pages they make are stored by a tiny link service so they can be shared and reopened free.
Type a thing (or tap a suggestion; Scissors is a built-in demo). Pick a language and who it's for (a curious adult, a 10-year-old, an engineering student).
Show me. Claude works out the parts and steps and writes the picture prompt (~1¢); OpenAI's image model paints the exploded view (~4¢ at high quality, set under More options). Both through OpenRouter. About 30 seconds.
One story in three chapters, with a bar pinned at the top (⚙️ How it works · 🏭 How it's made · 💰 What it costs) that fills as you scroll and jumps to a chapter. How it works: on wide screens the picture stays put while the numbered parts and steps scroll past; it starts whole and comes apart as the parts come up (a switch flips it by hand), and each part's number pulses as its card passes. Then the real thing in Wikipedia photos and a Sketchfab model to spin, which loads only when you reach it. Each part has a "How does this work?" link. The words appear as soon as Claude answers (~5 s); the pictures fill in as they are painted. The How it's made and What it costs chapters open with a title card and load their code and content just before you reach them.
Share. A short link (…/#scissors/s=k7x2mq9) opens the same page for anyone, no sign-in. Download report saves one self-contained HTML file with the picture inside; Download picture saves the image. Paint again keeps the words and paints new pictures; Remake starts over with new words too.
🏭 Factory comes in three versions, each its own plan: ✋ hand-made (labour-heavy, simple tools), 🤝 mixed and 🤖 automated (robots; people only where they are actually needed). A switcher picks one; once two are planned a table compares people, automation, floor space, output, investment and cost to make one. What it would take to make the thing, in the order it happens: raw materials, which parts are made and which are bought, how each is made, assembly, quality control, packing and shipping. Every step is marked 🤖 machine, ✋ people or 🤝 both, with the reason. A line diagram and an automation bar sum it up; a floor plan drawn to scale shows each zone's size and the machines' footprints, with a painted overview of the floor above it. One click, ~12¢ for the first version (Claude ~2¢, the floor ~4¢, the before/after pictures ~6¢ at low quality, six at a time) and less for the others, which reuse every before/after picture that shows the same thing; the scissors demo has one built in. Under the plan, before → after pictures show each zone's work as it comes in and as it leaves (a coil of steel → stamped blanks → hardened blanks…), made with the plan (an older plan gets a button): Claude describes each state, the image model paints each one as a small square, reusing a picture when one zone's input is the last zone's output. It is saved with the page, so share links carry it.
💰 Money. Approximate financials for the three setups side by side, for a region you pick: land and building, people and monthly salaries, equipment, power, other running costs and materials (Claude estimates the inputs, ~2¢; the page does every sum). Every input can be changed — a few overall controls (output, salaries, power price, materials, equipment and running costs) and every line in the detail — and the page recalculates at once; numbers you type are exact, the AI's are shown as ranges. A markup slider shows the selling price, monthly profit and how long each setup takes to pay back, against the typical market price, and the markup needed to pay back in three years.
🌍 Where it comes from, and where to make it, at the top of the What it costs chapter: a world map with arrows from where each raw material usually comes from to the factory, and every likely country shaded by what it would cost to make one there (hand-made, mixed or automated). Claude estimates the sources and each country's factory wage, power price and rent once (~2¢); the page scales the money estimate, including your own numbers, by them. Tap a country for its figures and a one-click full money estimate there, which moves the factory.
Right-to-left languages (Arabic, Urdu, Persian) flip the layout and switch to Noto Naskh Arabic. The home page shows a page from the gallery under the headline and a Browse answers grid, free to open.
Question + language + audience
├─ 1. Claude Sonnet via OpenRouter (strict JSON): title, hook, kind, 4–8 parts (name, English name, role),
│ 3–6 steps, fun fact, Wikipedia title, 3D search terms, and an image prompt that lists the parts
│ in order as numbered badges
├─ 2. openai/gpt-image-2.5-flare, via OpenRouter's /api/v1/images, paints the 1536×1024 picture (b64 → jpeg)
├─ 3. The page (words + picture) is POSTed to the link service, which stores the picture separately
│ and answers with a 7-character id; the browser keeps only the id
├─ Wikipedia (via the service, cached) → photos of the real thing
└─ Sketchfab (via the service, cached) → embedded 3D viewer
#flags.#admin=<ADMIN_SECRET>) gets an Add to gallery button.worker/worker.js; an older link service drops the plan and the page keeps it in the browser only.svg-era. It produced lively but unrecognisable drawings; a painted picture wins.| File | What |
|---|---|
index.html | the page |
style.css | picture-encyclopedia look (cream paper, ink lines), dark mode, RTL, mobile |
app.js | sign-in and keys, the prompt and schema, painting, storing, sharing, gallery, Wikipedia and Sketchfab |
i18n.js, i18n/*.js | the page's own labels: English in i18n.js, each other language in its own file, loaded only when a page in it opens; tx('key') picks the page's language |
scripts/make-gallery.mjs, .github/workflows/gallery.yml | makes complete gallery pages (page, three factory versions, money) on the live site; run from the Actions tab |
atlas.js, world.json | the map: its prompt, the per-country sums and the drawing; country outlines from Natural Earth via world-atlas, projected and keyed by country code (85 KB, loaded with the chapter) |
money.js | the What it costs chapter: the estimate's prompt and schema, the sums, cards, payback and cost bars |
factory.js | the How it's made chapter: the plan's prompt and schema, the line diagram, the to-scale floor plan, the factory part of the report |
demo.js, demo-scissors.jpg | the built-in demo page, with its factory |
worker/worker.js | the link service: stores pages and pictures (each picture written once; pages point to them, so a save costs one write plus one per new picture — Cloudflare's free plan allows 1,000 writes a day), short links, gallery, flags, Wikipedia and Sketchfab lookups (Cloudflare Worker + KV) |
logo.svg, favicon-48.png, apple-touch-icon.png | gear-with-a-question-mark badge, also the favicon and home-screen icon |
og-image.jpg, robots.txt, sitemap.xml | link preview picture and search-engine files |
factory.js and money.js load just before the reader scrolls to (or jumps to) the How it's made chapter. Every AI request sends its fixed instructions as a cached system prompt (repeats within a few minutes cost a tenth) and the product, language and reader in the user message.
| What | Where it goes |
|---|---|
| Your question and language | Through OpenRouter to Claude and to OpenAI's image model when you press Show me; to Wikipedia and Sketchfab to find photos and models |
| The page you get | Stored by the link service so it can be shared and reopened; it holds no name or account |
| OpenRouter sign-in | Your browser, until you sign out |
| Analytics | Page views, and which answers get opened, counted by GoatCounter (no cookies, nothing personal) |
MIT. Photos and 3D models belong to their authors under their own licences.
scripts/make-gallery.mjs opens the live site in a headless browser and, for each thing, makes the page, plans all three factory versions, estimates the money and adds the result to the gallery — about 30¢ and 5–10 minutes per thing. To run it on GitHub: add the repository secrets OPENROUTER_KEY and HOW_ADMIN_SECRET, then Actions → Make gallery pages → Run workflow (it stops starting new things once max_spend dollars are spent). The link service must be running the current worker/worker.js, or the factory and money parts are not kept.