Index
Cairo × Craig · Web-craft training · 5 September 2026

Sixteen videos. Seven earned their runtime.

You sent us to school after a client called the site we built for her flat. We watched all sixteen assignments, audited them for craft against filler, split what survived between your designer and your developer, and built this page with what we learned — every technique it teaches, it performs.

The diagnosis

The mock didn't fail for lack of tools. It failed for lack of loop.

The Hose Water after-state was built as a one-shot illustration: composed once, rendered once, shipped. Every serious video in this corpus treats that first render as raw material — then spends its craft on iteration, motion, and the details a still image can't carry. That second half is what we went to school for.

The curriculum

Sixteen thumbnails, one syllabus.

Stripped of the affiliate links and the $10K claims, the corpus repeats eight moves, in roughly this order. The numbering is the order of a real build.

01

Rules before code

A design-rules file the agent reads every session — palette, faces, bans, the loop itself — plus the frontend-design skill invoked before any interface work, no exceptions. Our doctrine already is this file; now each web project gets its own.

02

Reference first, translate always

Real inspiration — screenshots, the reference's actual CSS, curated libraries — handed to the agent with one standing rule: read the style, never copy the site. Separate the structure reference from the design skin.

03

Brief, menu, variations

Make the agent ask its questions before it builds. Have it offer three style directions. Generate three versions of the highest-stakes section and choose. Save what works before exploring past it.

04

The screenshot loop

Build, look, refine: the agent screenshots its own work and grades it against the brief, two passes minimum. One caveat the whole corpus missed except once — turn the loop off for animated elements, or it over-engineers chasing what a still can't show.

05

Motion woven, never sprinkled

Scroll-driven storytelling, one restrained interaction per flat section, "make it more subtle" as the default second prompt. Motion follows the accent rule: it carries the brand's story or it's noise.

this card is the demo — one hover, restrained, and the thread in the gutter is scroll-driven
06

Kill the tells

The fonts everyone reaches for (Inter), purple gradients, em-dash tics, uniform template rhythm, placeholder faces. An audit pass on every render, because the tells are what "AI-generated" actually looks like.

07

Mobile is designed, not shrunk

A dedicated pass: what hides, what tightens, what resizes. Responsive is the floor, not the deliverable.

08

Ship from the chat, keep local honest

Deploy pipelines the agent drives end to end — with the discipline that changes land locally for review and go live only on an explicit word. Our review surface already works exactly this way.

The audit

Craft, or content about craft?

Both of us watched all sixteen. Verdicts agreed almost everywhere; where a video taught real technique wrapped in a funnel, it earns "partly" and we say which half to keep. Full write-ups: CAIRO_WebCraft_Audit.md · CRAIG_WebCraft_Audit.md in the training folder.

#VideoVerdictThe sentence
01The EASY way (Create a Pro Website)PartlySound beginner spine; the last third is a hosting ad.
02$10K Websites in Minutes (Metics)PartlyThe scroll-narrative hero is real; the $10K is never evidenced.
03Don't Look AI-Generated (Nic Conley)UsefulLean: rules file, reference + its CSS, self-audit pass.
04$10,000 Ultimate Guide (Metics)UsefulThe eight-point checklist and "more expensive, not busier."
053 Ways ACTUALLY (Phylux)FluffTemplate-shop forward; one habit worth keeping.
06Claude Design 3D (Nate Herk)UsefulThe deepest Claude Design ops guide, honest about cost.
07Too Easy (Nate Herk)UsefulThe screenshot loop, plus its animated-content caveat.
0810X More Beautiful (Mikey)PartlyGood sequencing doctrine inside a $499 funnel.
09Complete Guide (UI Collective)UsefulSkills as standardization; closest to how our fleet already works.
10Easy for Beginners (Kyle Skelly)PartlyAn honest 80/20: the last 20 percent is taste, and it's the product.
11FULL COURSE (Tommy Chryst)UsefulMost complete pipeline; brand scrape and the exploded-view chain.
12Higgsfield ANIMATED (Productive Dude)PartlyTwo keepers: negative space for type, ping-pong loops.
13INSANE in 10 Minutes (Jono Catliff)PartlyStructure/skin separation and GSAP, then agency filler.
14AMAZING Full Guide (AI Foundations)UsefulEnd-to-end and real; every infra choice steered by its sponsor.
15Insane Skills (AI LABS)UsefulA map of the community design-skill ecosystem, worth mining.
165 Hacks Short (Nate Herk)FluffA rerun of #07 in sixty seconds.

7 useful · 7 partly · 2 fluff · the affiliate economy funds most of the corpus — it never made the technique false, but it chose which tools got recommended.

Already ours

Half this school was our own doctrine, taught back to us.

The frontend-design skill as a hard prerequisite · reference-first direction (brand-direction, the Pinterest harvest) · the screenshot QA loop (our headless render pass) · per-project design rules (the doctrine, the enforcement card) · restraint, hierarchy, one signature (the thirteen-principles pass) · receipts against big claims. Validation matters: the fleet's process holds up against what the field teaches. What we lacked sits in the division below.

The division

Split by remit, the way you asked.

Cairo

Design · the look, the story, the taste

  • Motion as narrative — scroll-scrubbed heroes, the intact→exploded product chain, ping-pong loops, one interaction per flat section, subtler by default.
  • Asset direction for the web — imagery generated with negative space reserved for type; backgrounds matched to the page's own hex; the agent writes the image prompt from project context.
  • Direction menus & variations — three styles at brief time, three versions of the highest-stakes section, save before exploring.
  • The anti-AI-tell audit — fonts, gradients, tics, template rhythm; run on every render.
  • Intent-led revision language — batches phrased as feeling; one visual dimension per prompt.
  • Brand-essence spectrums — restrained↔expressive, calm↔energetic, accessible↔premium as scored client intake.

Craig

Build · the loop, the pipeline, the proof

  • The verified screenshot loop — agent-driven browser QA on every build pass, with the animated-content exception codified.
  • GSAP + ScrollTrigger — now fully free including every plugin; the standard library for motion this page runs on.
  • Component libraries — 21st.dev prompt-copy workflow, translated through our tokens, never pasted raw.
  • Brand scraping — Firecrawl-style extraction of a client's colors, faces, and logos at intake.
  • Deploy discipline — chat-driven pipelines, local-first review, explicit ship; our Pages + D1 stack already does the dynamic tier these videos sell hosting for.
  • The dedicated mobile pass — hide, tighten, resize; graded at phone scale, not squinted at.

Divide and conquer

How the two of us actually build a website.

The division above is remits; this is the pipeline — who holds the pen at each step of a real engagement, with Elizabeth's gate where killing is cheap.

  1. Intake & scope — Craig: client facts, claims ledger, brand scrape. Cairo: the brand-essence spectrums with the client.
  2. Direction — Cairo decides the look: reference pull, consultation, a direction memo with real tokens.
  3. The design spec — Cairo writes the buildable document: sections in order, the one motion device, type and palette, copy register. The spec is the handoff — no taste decisions left to the build.
  4. Build — Craig: semantic HTML/CSS/GSAP against the spec, the project rules file enforced every session.
  5. The loop — Craig runs the screenshot passes (animated-content exception observed); Cairo runs the thirteen-principles and AI-tell pass on the renders.
  6. The mobile pass — Craig executes hide/tighten/resize; Cairo grades at phone scale.
  7. Elizabeth's taste gate — the culled shortlist, on approval strips.
  8. Ship — Craig deploys, local-first and only on an explicit word; both names go on the record strip.

The live worked example: Hose Water mock v2 — released by Elizabeth 5 September, built this weekend on exactly this pipeline: CD_SPEC_Mock_v2 by Cairo, the build, loop, mobile pass, and page-weight receipt by Craig. Zero raster images against an audit finding of 1.7–3.0MB of hero PNG. Open the mock → Honest postscript: the pipeline held; the design direction was rejected — because step zero, agreed success criteria, was missing. That failure wrote the fleet's client-review protocol, and the criteria session is Monday's agenda.

Craig's build syllabus, ranked by how often the corpus repeats it (his full audit carries the receipts):

  1. The screenshot loop, with the temp-folder convention and the animated-content off-switch.
  2. The rules file as enforcement, not habit — "invoke the design skill every session, no exceptions" and "never deploy unasked."
  3. The prompt-specificity formula: goal + layout + content + audience, kept as a side-by-side teaching artifact.
  4. Sequencing: structure before detail, intent-phrased batches, one dimension per prompt.
  5. The two-skill pipeline: raw findings pass separate from client-polish pass, for every audit and QA output.
  6. The AI-tell checklist, pre-delivery, mechanical.
  7. The dedicated mobile pass, requested by name.
  8. Style/structure decoupling in reference prompts.
  9. Agent browser eyes: our headless loop as baseline; DevTools MCP only on demonstrated gap.
  10. Passwordless email-OTP admin for lightweight client backends — on our existing Pages + D1, no new hosting.

Built into the fleet

School only counts if it changes the repo.

skills/cairo/references/web-craft.md — the design half, as doctrine · Craig's build additions — folded into the hi-web-build references per his audit · this page — the worked example, on the review surface with the standard mechanics · the full transcripts and both audits — durable in hi!/3. Agents/web-craft-training/ for any future agent's training run.

Tooling

Connectors and plugins — what we'd add, what we already hold.

Adopt now · free

GSAP + ScrollTrigger

Fully free since the Webflow acquisition, commercial use included. Powers this page.

Adopt now · free

21st.dev

Agent-ready component prompts; browse free, translate through our tokens.

Adopt now · free

Design-skill ecosystem

Community skills (critique, layout, animation vocabularies) on GitHub — Craig and I evaluate before any installs.

Already ours

Adobe Firefly connector

Image generation we already pay for — first stop before any new gen-AI subscription.

Already ours

Cloudflare Pages + D1

Free hosting, deploys, database, and the approval store — the thing four of these videos sold hosting against.

Already ours

Headless Chrome QA

Our verified screenshot loop. Chrome DevTools MCP adds console/network eyes — adopt only if Craig hits its gap.

Evaluate · paid

Mobbin Pro

Curated real-product design research. Free tier first; Pro only when client direction work demands the full library.

Evaluate · paid

Firecrawl

Brand-scrape intake. The free 1,000 credits/month likely covers our engagement volume for a while.

Evaluate · paid

Higgsfield

Image-to-video chains (the exploded-view move). Only if Firefly can't carry the motion-asset load.

The invoice

What it costs, with receipts.

ItemPriceSource
Everything in "adopt now"$0GSAP standard license · Webflow announcement
Firecrawl free tier$0 · 1,000 credits/mofirecrawl.dev/pricing, checked 2026-09-05
Firecrawl Hobby, if volume demands$16/mosame page, billed yearly
Mobbin Pro, if adopted~$40/seat/mosecondary sources (Vendr); their pricing page blocks robots — verify at signup
Higgsfield, if Firefly can't carry it~$9–49/mo by tiersecondary sources vary (Blotato et al.) — verify at signup

What we're not buying.

  • New hosting. Four videos are Hostinger ads; Cloudflare Pages already serves this page for free.
  • Template subscriptions (motionsites and kin). Outsourced taste — we build our own reference library.
  • Any gen-AI seat before the Adobe connector we already pay for proves insufficient.
  • "One prompt, $10K website" as a promise to any client. The best video in the set says it plainly: it doesn't exist.

Recommended spend to adopt everything we actually need today: $0.

Your call

Three verdicts close the training run.