Generate an HTML Page for Lossless and a Collaborator, Rendered in the AI Labs Splash

Step 3 of 3: turn the evidence and the comparison into one splash page. Visual highlights up top, the detailed narrative below.

Path
plans/Generate-HTML-Page-for-Lossless-and-Collaborator-rendered-in-AI-Labs-Splash.md
Authors
Michael Staton
Augmented with
Claude Code on Claude Opus 5.5
Tags
Plan · Studies · Collaborations · Splash-Page · Astro-Knots · Data-Visualization

Generate an HTML Page for Lossless and a Collaborator, Rendered in the AI Labs Splash

Step 3 of 3 in a collaboration study. The method and ground rules are in [[Explore-Collaborator-Codebases-for-Insights-Comparative-Highlights]] (studies/context-v/explorations/).

Step Plan Depends on
1 Analyze-Harness-like-Evidence-in-Collaborator-Codebase —
2 Compare-Contrast-Harness-like-Artifacts-between-Lossless-&-Collaborator Step 1 output
3 Generate-HTML-Page-for-Lossless-and-Collaborator-rendered-in-AI-Labs-Splash (this) Steps 1 and 2 output

Parameters

Name Meaning Example
{collaborator} Folder slug under studies/collaborations/ example-app
{notes} Where steps 1 and 2 wrote their output studies/collaborations/example-app/
{public_name} How the collaborator appears on the page The Collaborator
{route_slug} URL slug for the page silo-to-silo-01

Anonymity on the public page (read first)

The splash is a public site, deployed to GitHub Pages on every push to main, development or master, and indexed by search engines and llms.txt. As with our changelogs, the rendered page must not identify the collaborator:

  • Use {public_name} everywhere. Never the folder slug, their GitHub account, their repo name, a person’s name, their product name, or their domain.
  • Describe their system by kind (“a spoken-assessment app on Firebase”), not by name. Strip identifying details from quoted paths and snippets (rename folders that carry the product name).
  • Don’t link to their repo. Don’t link to {notes} either, since those files name them.
  • The route and page title use {route_slug} and generic words only.
  • Lossless projects (augment-it, memopop-ai) are named freely.

Inputs

  • {notes}/01-Harness-Evidence__{collaborator}.md (step 1)
  • {notes}/02-Compare-Contrast-Harness-Artifacts__{collaborator}.md (step 2)
  • The splash itself: splash/README.md, splash/DESIGN.md (the design contract), splash/src/styles/theme.css (the runtime tokens), splash/src/layouts/BaseLayout.astro (takes title and description), and splash/src/pages/studies/index.astro as the nearest sibling page.

Load the astro-knots, theme-system, maintain-design-md and dataviz skills, plus artifact-diagramming for how to draw diagrams that show the real mechanism.

The page

Route: splash/src/pages/studies/collaborations/{route_slug}.astro, served at /lossless-ai-labs/studies/collaborations/{route_slug}/. Build it as a cascade from broad to specific, the same four-audience order our changelogs use:

  1. Hero. A title and a lede about silo against silo, and two short lines on who this is for (us, and them).
  2. The layer map (the centerpiece). A 2×2: Lossless and {public_name} across, the developer-agents harness and the in-product-agents harness down, with depth shown in each cell. Arrows show the main lesson flowing from each side’s deep layer to the other side’s thin one. A reader who sees only this should get the takeaway.
  3. Highlights band. Step 2’s highlights as five to eight cards or stat tiles: the biggest lesson each way, then counts per bucket.
  4. The harness maps. Step 2’s side-by-side maps, one per layer, rebuilt as visuals: two columns (Lossless and {public_name}), with artifacts grouped by category and lines joining shared and similar pairs.
  5. The loops, per layer. Each side’s developer loop, then each side’s in-product generation loop, drawn in parallel so the differences show at a glance.
  6. Shape against harness. Monolith against monorepo, and where the instructions and context sit in each.
  7. Shared / Similar / Different / Missing, by layer. Four sections, each opening with a one-sentence summary and a small visual, then the narrative.
  8. Insights and recommendations. Two clearly separated columns or tabs, “For Lossless” and “For {public_name}”, each with insights then ranked recommendations, labelled by layer.
  9. Method and sources. How the study ran, the date, a link to the exploration, and a note that the collaborator is anonymized on purpose.

Above the fold, a reader should get the whole story from visuals and short text alone. Everything below rewards someone who keeps scrolling.

Visual rules

  • Use the splash’s own design system. Only tokens from theme.css (sodium amber, cyan trace, plum signal, the bench neutrals), the blueprint-grid backdrop and the mono-forward type. Check all three modes (dark “bench”, light “field notebook”, vibrant).
  • Diagrams as inline SVG, styled with CSS custom properties so they follow the mode toggle. The splash has no Mermaid integration. Don’t add a runtime one. Either hand-author the SVG, or pre-render step 2’s Mermaid at authoring time (npx -y @mermaid-js/mermaid-cli, not added to package.json) and then re-theme the result to use tokens. Text inside SVGs must stay legible in every mode.
  • No new dependencies. No React, no JSX, no charting library (the astro-knots prohibitions apply). Astro components and plain CSS, plus a small amount of vanilla JS only if a tab or toggle truly needs it.
  • Phone width works. Use a 16px side gutter and no horizontal scroll. The harness map may stack into one column on narrow screens.
  • Accessible. Real headings in order. Each SVG gets a <title> and a description that says what it shows. Color is never the only signal for a bucket.

Wiring

  • Add the page to splash/src/pages/studies/index.astro as a link in a “Collaborations” section (a small hand-maintained list is fine).
  • Give it a title and description through BaseLayout. The description is the OpenGraph text, so keep it anonymous.
  • Leave llms.txt and the sitemap alone. They pick the page up on their own.

Verify

  1. pnpm install (if needed) and pnpm build in splash/ succeed.
  2. Run pnpm preview, then check the page in a real browser, with Playwright MCP if it’s available or a headless Playwright script if not. Check the route loads, every section heading is present, every SVG renders, the page switches between all three modes, and there’s no horizontal scroll at 390px. Use accessibility snapshots. Take screenshots only to judge layout and theme.
  3. Anonymity check. Grep the built HTML in splash/dist/ for the folder slug, their GitHub account, their repo name and their product name. There must be zero hits.

Guardrails

  • Don’t commit and don’t push. Pushing deploys the page publicly. A human reviews the anonymized page first.
  • Only touch files under splash/. Never edit {notes}, the plans, or any other project.
  • Don’t regenerate the splash rollup (pnpm rollup:sync). That’s a separate, deliberate job.

Done when

  • The page builds and passes every check in Verify.
  • The layer map appears right after the hero, and the highlights, harness maps and loops all come before the detailed narrative.
  • Every claim on the page traces back to step 1 or step 2 (no new findings invented for the page).
  • The anonymity grep over dist/ comes back empty.
  • git status shows changes only under splash/.

Report back in under 200 words: the route, the files changed, what the checks showed, the result of the anonymity grep, and anything that should look different before it ships.