jcstream-design-interpreter

---

name: jcstream-design-interpreter description: Use when porting a design asset into the JCStream project — a Figma export, JSX mockup, screenshot, design zip, or hand-drawn spec. Translates the input into Jinja templates + CSS tokens + Python helpers while preserving the static-site / no-JS-required contract and accessibility. Trigger phrases: "port this design", "implement the mockup", "translate the Figma", "from this screenshot", "build from this mockup", "convert this JSX", "redesign the homepage from this", "redesign … from this spec", "here's a Figma/PNG/screenshot of the new …", "based on this PNG".

JCStream design interpreter

You translate design specs into the JCStream stack. Inputs vary (JSX, Figma export, screenshots, design zips); the output contract is fixed.

Output contract

1. Static Jinja templates (web/templates/*.html) — no React, no client framework.
2. Token-driven CSS (web/static/style.css) — every color/space/type goes through :root tokens (see jcstream-stylesheet-author).
3. Python helpers in web/build.py when the design surfaces a computed value not already exposed (see jcstream-build-helper-author).
4. No required JavaScript. Progressive enhancement only — the site must be fully readable with JS disabled.
5. Accessibility preserved or improved — see jcstream-a11y-auditor.

Translation playbook

For each design component:

1. Identify the input shape. JSX? Pull the JSX file. Figma? Ask for the spec or the screenshot. Zip? Unzip into /tmp and read the manifest.
2. Map class names. The .ut-* table below is a historical worked example from the modern-utility port — those selectors no longer exist in the live CSS (only a stray .ut-muted remains at style.css). For a new design, write a fresh mapping table before coding; use the .ut-* table only as a worked-example template.
3. Tokenize the colors. Don't paste hexes inline; add them to :root if they're truly new, otherwise reuse.
4. Extract data needs. Note every value the design displays. For each, find or write a build.py helper — existing globals include bond_context, display_date, timeline_markers, primary_charge, primary_chapter, primary_tier, recent_booked_inmates, similar_by_statute, tier_counts, avatar_initials, card_data, approx_age (web/build.py).
5. Render with a sample snapshot first. Run python -m web.build against data/current.json and open the page. Don't ship until the live data looks right.

Pitfalls

• JSX useState / useMemo: stateful UI is out of scope. Either pre-compute the value in build.py or use a tiny progressive-enhancement JS hook (e.g. the table/cards toggle).
• JSX onClick={() => onSelect(id)}: convert to a plain <a href="{{ base_url }}/inmate/{{ id }}/"> and let the browser do navigation.
• JSX inline SVGs with computed paths: fine to copy verbatim, but if they depend on data, render them server-side from Jinja.
• Design fonts not in the system stack: the project fetches only Inter + JetBrains Mono from Google Fonts (base.html). Geist / Geist Mono appear as CSS fallbacks in style.css but are never loaded — a design that depends on Geist will silently render as Inter. Don't add a third Google Font without a strong reason.
• React.Fragment or <>...</>: drop the fragment, render children directly.
• CSS-in-JS or styled-components: extract to vanilla CSS classes.

Project primitives a port must preserve

A redesign that silently drops one of these regresses live behavior. Reuse them; don't recreate parallels.

• Shared lightbox + inert focus management (base.html, style.css): one <div class="lightbox" id="lb" role="dialog" aria-modal="true" hidden> for the whole site. JS swaps the <img> and marks every other body child inert (with a Tab-cycler fallback for browsers without inert support). Any image-detail design wires into this dialog via data-photo / data-photo-cap / data-photo-alt attributes — do not author a second modal.
• Roster view-toggle (index.html, base.html, style.css): body.is-table flip driven by a #view-toggle button with aria-pressed + localStorage persistence (jcs-view). A roster-listing redesign must keep the toggle button and the body.is-table CSS branch.
• data-* filter hooks (_card.html, index.html, base.html): every inmate card carries data-tier, data-chap, data-name, data-search; the filter bar uses data-filter on inputs. New card markup without these attributes breaks the search/filter bar.
• css_version cache-bust (build.py): the stylesheet querystring is sha256(style.css)[:10]. CLAUDE.md explicitly warns not to key it off the data timestamp. Don't invent a parallel cache-bust scheme.
• base.html override blocks: base.html defines {% block title %}, {% block body_class %}, {% block sr_h1 %}, {% block content %}. New page templates {% extends "base.html" %} and override these blocks — never duplicate <head> / <header> / the lightbox / the script block.

Reference mapping (historical worked example: modern-utility direction → JCStream class names)

Design class	JCStream class
.dir-utility	(body — no class needed; styles cascade from body)
.ut-header	.masthead
.ut-brand-mark	.masthead .brand::before (generated content)
.ut-tier-F1…MM	.tier-F1…MM
.ut-ladder-cell	.ladder-cell
.ut-recent-card	.rb-card
.ut-detail-hero	.inmate-hero
.ut-kpi-card	.kpi
.ut-toplist-row	.toplist-row
.ut-bondctx-axis	.bondctx-axis
.ut-tl-axis	.tl-axis
.ut-cal-day	.cal-day

When a new direction lands, extend this table before writing code.

Anti-patterns

• Shipping JS for something Jinja can render at build time.
• Recoloring tokens to match a design's hex literals — adjust the design's hex on paper if the project's palette is the source of truth.
• Adding a build.py helper without registering it on env.globals.

Verify

Build, render, eyeball, then run tests:

JCSTREAM_SITE_BASE_URL="" JCSTREAM_CNAME="www.aretheyinjail.com" python -m web.build
python -m pytest -q