…p teal in Testimonials

- ComparisonSection: match the standard card chrome
  (border-white/[0.06] bg-white/[0.012], no shadow, no teal glow),
  add the static dot-grid backdrop, replace Lucide check/X with
  monochrome dot-vs-dash marks, give the Refactron column a
  continuous vertical accent via border-l-white/[0.22].
- QuickstartSection: same chrome alignment, dot-grid backdrop,
  emerald command text -> white, '$' prompt -> faint chevron.
- TestimonialsSection: keep the cream paper inversion as the
  intentional editorial beat; strip the lone teal radial in the
  backdrop and teal chart bars to neutral grays (only color we've
  removed everywhere else on the site).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

- Hero leads with "Every codebase has code nobody wants to touch"
  (no eyebrow above the headline); a two-line subhead carries the
  promise, signed off with a mono founder signature.

- Founder card restructured so "the moment" (the hackathon scene
  that started Refactron) is the lede, not a buried paragraph.
  Identity block is now two mono fields (role / building); receipts
  row and X + LinkedIn affordances kept.

- Portrait chrome simplified: single ring + soft shadow, no more
  dual outer glow + ring stack.

- Card chrome unified with Workflow / WhatWeDo / Comparison /
  Quickstart: border-white/[0.06] + bg-white/[0.012] + static
  dot-grid backdrop. Dropped panelSurface + panelInnerGlow.

- Severity badges in the health digest stripped of rose/amber/sky
  colors. SeverityTag is a single monochrome chip where dot alpha
  and label brightness encode severity (ERR / WARN / INFO).

- Approach step badges flattened to small rounded squares with
  two-digit padded numbers; ListMarker hairline dash replaces the
  colored check-in-circle bullets.

- "Why I'm Building" rebuilt as a letter close — A LETTER eyebrow,
  short pull-quote with white emphasis on the key phrases, hairline
  rule + mono signature. Dropped the giant decorative quote mark.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

The three middle sections were the weakest beat on the page —
each one was a generic card with a text list inside. Replacing
each with a visual register a developer actually recognizes:

- The Problem -> a real terminal-output card. macOS-style dot
  strip, prompt, command, ✓ scanned line, then five log lines
  with bracketed [ERR ] / [WARN] / [INFO] tokens whose
  brightness encodes severity. Footer shows a count breakdown.
  SeverityToken inline component replaces the boxed SeverityTag.

- My Approach -> CI/pipeline visual. Top: a 5-dot horizontal
  overview where dot brightness fades through the pipeline.
  Below: vertical pipeline with a gradient connector line
  threading the step circles, each step as a numbered black
  circle (01..05) with title + detail.

- What Safe Means -> each card now leads with a per-constraint
  SVG diagram (SafetyVisual): READ|WRITE state for 01, the
  PROPOSED -> APPROVED flow with a check for 02, a stacked
  syntax/tests/invariants pass list for 03, five progress nodes
  decreasing in brightness for 04, an UNDO loop arrow from "now"
  back to "prev" for 05. Outcome label is now a chip; outcome
  body bottom-anchored so card heights look planned.

All monochrome. All static after mount. The static dot-grid
backdrop on each card stays.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Hero illustration is now a tactile, dark-glass "REFACTRON ·
ENGINE" tile — built in CSS, not SVG hairlines, so it actually
reads as a piece of hardware rather than a blueprint drawing.

Surface treatment:
  - Aspect-square rounded-[28px] tile, soft top-down satin
    gradient + radial highlight at the top, inset upper rim
    bright + lower rim shadow, large drop shadow underneath
    for elevation. Subtle outer halo (white at very low alpha).
  - A 1px top-edge highlight hairline across the top inside the
    tile sells the "milled edge" look.

Composition inside the tile:
  - Top-right: a small monochrome "R" sigil glyph.
  - Centre: a 9x9 chip-die dot matrix. Dot alpha falls off
    radially from the centre; the centre dot pulses gently
    via a single CSS keyframe (rfn-die-pulse, ~2.6s loop) —
    "powered on" feel. A faint radial halo sits behind the
    matrix.
  - Bottom-left: REFACTRON / ENGINE in mono.
  - Bottom-right: a row of 5 status LEDs; the leftmost shares
    the same pulse animation as the core ("active" state).

Periphery (outside the tile):
  - FIG · 01 at the top-left.
  - DETERMINISTIC · V0.5 at the bottom-right.
  - At xl+, a vertical rail label running down the right edge:
    ANALYZE · REFACTOR · VERIFY · DOCUMENT.

Removed: HeroEngineSchematic + Gear + GearProps (SVG line-art
version). Replaced with HeroChipModule + ChipCoreDie. No iso
primitives. Pure white at varied alpha on near-black surfaces.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Replaced the flat-SVG SafetyVisual line drawings with five CSS
glassmorphic widgets — same satin / inset-shadow / soft-glow
language as the hero chip module, scaled down. Each one is a
small "control surface" element with one iconic affordance.

Shared base recipes (defined once, reused per card):
  - glassBase  — neutral satin tile with inset rim shadow.
  - glassActive — brighter top-down gradient, raised rim,
                  outer ambient glow.
  - glassDim   — dim background with dashed border (proposed /
                 inactive state).

01 SafetyReadOnly — two pills: an active READ glass pill (raised,
   glowing dot) and a dim WRITE pill with a horizontal strike
   across it; dashed hairline between them.

02 SafetyApprove — stacked PROPOSED → APPROVED mini-cards.
   PROPOSED uses the dim/dashed glass style. APPROVED uses the
   active/raised glass style and carries a check glyph.

03 SafetyVerify — a tile with three glowing LED rows (syntax /
   tests / invariants), each with a soft white shadow and a
   right-aligned "PASS" tag.

04 SafetyIncremental — five small "key cap" tiles in a row.
   Each has its own subtle satin gradient + rim shadow.
   Active key is brighter with an ambient glow; the rest fade.
   Two-digit padded numbers inside (01..05).

05 SafetyRollback — three round glass buttons in a row over a
   dashed arc trail: a dim "prev" indicator on the left, a large
   central rollback button with a radial-gradient surface and an
   embossed undo arrow, and a glowing "now" indicator on the
   right.

Same monochrome palette. No iso primitives. Pure CSS — no extra
SVG hairline diagrams.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

The inline SVG I hand-rolled for the central rollback button was
mis-pathed (the arrow-hook and the arc didn't visually join up,
so it read as a backwards "C" with a stray tick rather than the
universal counter-clockwise undo glyph).

Swapped it for Lucide's `RotateCcw` — the canonical rollback /
undo affordance. Stroke width tuned slightly heavier (1.75) so
it reads cleanly against the dark glass disc.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Built a deliberately beautiful ASCII pattern backdrop and applied
it to the two sections that benefit most from atmospheric mood:
the hero (the first impression) and the "A letter" close (the
emotional payoff).

Implementation:

  - buildAsciiPattern() — deterministic generator. Walks an
    NxM grid of cells; for each, picks SPACE or a character from
    a passed-in charset using a seeded LCG. Per-cell placement
    probability scales with distance from the centre (centerSparsity
    and edgeDensity), so density rises toward the edges and the
    middle stays open.

  - Two patterns are built once at module load:
      ASCII_MIST  — 180x80 grid of punctuation glyphs
                    (· . : , ; - _ = / < > ( ) [ ] { } |)
                    at very low alpha (rgba 0.05).
      ASCII_STARS — same grid dimensions, very sparse,
                    drawn from a brighter charset
                    (0 1 + * ✦ ░ ▒) at higher alpha (0.13).

  - <AsciiBackdrop /> stacks both layers as <pre> blocks with the
    correct font-family, line-height, and letter-spacing. A
    radial-gradient mask fades both layers toward a configurable
    centre point (maskAt prop) so foreground text stays crisp.
    A single CSS keyframe (rfn-ascii-breath, 9s ease-in-out) lifts
    the opacity slightly; the two layers run with offset delays so
    they never breathe in lockstep.

Placement:
  - Hero  — maskAt="32% 50%" so the soft centre falls under the
            headline column on the left, ASCII fills the right.
  - "Why I'm building" letter card — maskAt="35% 55%", same idea,
            ASCII surrounds the letter text without smothering it.

DotGridBackdrop is still the default everywhere else. Pure CSS,
no SVG, no canvas, no runtime regeneration after mount.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…section

Built a second ASCII backdrop that's figurative rather than
ambient scatter — two tapering "arm" silhouettes that sweep in
from opposite corners and almost touch in the centre (the
Creation-of-Adam composition, the spark gap left intact).

Generator (deterministic, module-level):

  HandBlade type — origin (ax,ay), tip (bx,by), thickStart,
                   thickEnd. Coordinates are in screen units;
                   y is multiplied by 2 at sample time to
                   compensate for monospace char aspect.

  HAND_BLADES — two blades:
    1) ax=-8, ay=78  → bx=78, by=46  (left arm, lower-left ⇒
       fingertip just left of centre)
    2) ax=168, ay=22 → bx=84, by=50  (right arm, upper-right ⇒
       fingertip just right of centre)
    Both taper from thickStart=9.5 at the wrist to thickEnd=0.6
    at the fingertip.

  bladeSD(c, r, b) — signed distance from (c, r*2) to the blade's
                     spine minus its lerped thickness. Negative
                     inside the blade, positive outside.

  For each cell I take the min SD over the two blades and bin it
  into a density palette:
      sd < -4   → ▓ (deep inside)
      sd < -2   → ▓/▒ random (transition band)
      sd < -0.5 → ▒
      sd < 0.4  → ░ (edge)
      sd < 1.5  → · or space (stray fringe)
      else      → space

  Final pattern is 160 cols × 48 rows, baked once into
  HANDS_PATTERN. The seeded LCG keeps it stable across renders.

<HandsAsciiBackdrop /> renders the pattern as a single <pre>
(font-size 13px, line-height 1.05, letter-spacing 0.06em, alpha
0.075). Radial mask peaks at the middle and fades to transparent
at the corners so it never fights with sectionFades.

Placed on the Founder + Origin Story section — the "hands
reaching, almost touching" metaphor lands well alongside the
hackathon-moment narrative without needing explanation.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…ll.txt

LLM/GenAI discoverability infrastructure. Three concrete changes:

- public/robots.txt: add an explicit AI-crawler block at the top
  of the file, ahead of the existing search-engine rules. Two
  groups, both with Allow: /
    • Retrieval bots (live AI search): OAI-SearchBot, ChatGPT-User,
      Claude-SearchBot, Claude-Web, PerplexityBot, Perplexity-User,
      Applebot-Extended, YouBot.
    • Training bots (long-term LLM memory): GPTBot, ClaudeBot,
      anthropic-ai, Google-Extended, Meta-ExternalAgent, CCBot,
      Bytespider, cohere-ai, cohere-training-data-crawler, Diffbot,
      FacebookBot.
  Refactron's goal is to be both indexed for live AI answers AND
  remembered in foundation models, so both groups are explicitly
  allowed (no more commented-out / allowed-by-default ambiguity).
  Also added Sitemap: line for /llms.txt alongside /sitemap.xml.

- public/llms.txt: new, hand-curated markdown site map per the
  llmstxt.org spec — H1 + blockquote positioning + Docs/Examples/
  Optional sections. Read by AI IDEs (Cursor, Claude Code, Windsurf)
  when a user points them at refactron.dev, and increasingly by AI
  search retrievers.

- public/llms-full.txt: new, auto-generated full LLM-ingestion doc.
  Concatenates (a) curated product overview + safety constraints +
  comparison table, (b) the origin story, (c) all FAQs from
  FAQSection.tsx, (d) every blog post (title + summary + full body)
  from posts.ts. ~56 KB, 872 lines, regenerates on every build.

- scripts/generate-llms-full.js: the generator. Mirrors the
  pattern of scripts/generate-sitemap.js — regex parsing of
  posts.ts and FAQSection.tsx, no ts-node required.

- package.json: prebuild now runs sitemap + llms-full back-to-back
  so the two stay in sync on every build.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…Application schema

LLM-targeted copy and structured data updates on the landing page.

Meta:
  - <meta name="description"> rewritten to lead with the phrases
    LLM users actually search for: "deterministic", "behavior-
    preserving", "no LLM in the loop", "verified before every
    write", "runs locally". ≤160 chars.
  - <meta name="keywords"> expanded to cover the high-intent LLM
    query phrases: "deterministic refactoring",
    "no-LLM code refactoring tool", "behavior-preserving",
    "legacy code modernization", "local-first developer tools".
  - og:description and twitter:description mirror the new meta
    description (same primary positioning across SERP previews).

Structured data (script type=application/ld+json):
  - SoftwareApplication: programmingLanguage promoted to an array
    (Python, TypeScript, JavaScript), softwareVersion added,
    free-tier Offer added, and a featureList of the seven core
    capabilities — each phrased as a fact AI assistants can quote.
  - NEW FAQPage block sourced from the six entries in
    src/components/FAQSection.tsx (does Refactron change code
    automatically / how is it safe / production-ready / where
    does code go / supported languages / open-source). FAQPage
    is the single highest-impact JSON-LD type for AI-assistant
    quoting in 2026.

application/ld+json blocks now = 4 (Organization, WebSite,
SoftwareApplication, FAQPage). Pre-render integrity preserved.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…ders

BlogPostPage:
  - Wired the existing useSEO hook so each post page sets its own
    <title>, meta description, keywords, canonical, og:title/desc,
    and twitter:title/desc from the matched post data. Falls back
    to a neutral "Post not found" set when slug doesn't resolve.
  - New useBlogPostingSchema hook: on mount it injects a
    <script type="application/ld+json"> BlogPosting block into
    <head> with headline, description, datePublished (ISO),
    author (Person with profile link), publisher (Organization
    with logo), mainEntityOfPage, url, keywords, inLanguage.
    Removes the script on unmount so navigation between posts
    keeps exactly one BlogPosting block live.

vercel.json:
  - New header rules for /llms.txt and /llms-full.txt mirroring
    the /sitemap.xml config:
      Content-Type: text/markdown; charset=utf-8
      Cache-Control: public, max-age=86400, stale-while-revalidate=43200
      X-Robots-Tag: all
    The X-Robots-Tag is the explicit signal to retrieval bots that
    these files are fair game.

End-to-end result: every blog post now ships with rich Article-
level structured data; both LLM index files are served with the
correct content type and a 24h cache.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>