Implementation notes: general /remix + carousel renderer (2026-10-11)
Running log kept by the implementing sub-agent. Decisions, dead ends, and what was verified and how.
Brief
Widen ~/.claude/skills/remix/SKILL.md from "Sanity Check issue to social posts" to "any piece to (a) LinkedIn carousel PDF, (b) X post, (c) newsletter section". Keep the Sanity Check mode. Build a carousel renderer (1080x1350, brand tokens, headless Chrome, one command). Run one real sample: Ship 30 Day 1 to a carousel. Founder approval: iMessage 2026-10-11 ("We definitely want our own brand guide, not Gamma"; "draw inspiration from those sources and roll our own skills").
Sources read (04:20 ET onward)
- [[2026-10-10-content-sdlc]] - /remix is the "Remix" entry point; carousel rendering was awaiting the founder's pick (brand HTML to PDF vs Gamma). His 2026-10-11 message settles it: brand HTML to PDF.
- [[CAPABILITIES]] - no carousel capability exists; index predates the Denominator pack (last updated 2026-05-23).
- Design contract:
02-sops/packs/design/denominator/(README, [[tokens]], [[ART]], [[components]]). The pack README says it supersedes [[DESIGN-denominator]] v3.1 for reading pages, so v3.1 (newsprint, hard shadows) is NOT used. - Carousel recipe: [[2026-10-05-linkedin-ai-classroom-map]]:121-123 (hook, re-hook, one point per slide, CTA; 5-16 slides; mid-roll CTA on long decks). Rewritten in our own words in SKILL.md.
- [[content-critic-rubric]], [[WRITING-rdco]], [[VOICE]], memory
feedback_x_voice_mismatch,feedback_no_em_dashes. - Source piece: [[2026-10-07-week1-drafts]] Day 1 (lines 23-45). Facts traced to [[2026-10-07-locked-topics-and-question-bank]]:19-27.
Decisions
- Design contract = Denominator pack v1 round 7 (paper
#F3F1EB, sheet, ink, one cobalt accent#2B57C7, 1px--linebox with 3px corners and no shadow, Source Serif 4 / Instrument Sans / JetBrains Mono). Field colours only inside the art frame ([[ART]] Recipe A). It is the founder's personal-brand contract and the newest one he ruled on (2026-10-06). [[DESIGN-denominator]] v3.1 is superseded for reading surfaces by the pack's own README. - No git / no PR.
~/.claude/skillsis not a git repo (git rev-parsefails), and the brief says no pushes. Edits are made in place. Original SKILL.md backed up to the session scratchpad before the rewrite. - Sanity Check mode kept verbatim except the "After Saving" step: it said "Report the remix to the founder via iMessage or Discord". Discord is retired (AGENTS.md) and the new skill never messages anyone, so that line now says the summary is handed back to the caller. Notion calendar update kept.
- Carousel pipeline = HTML, then Chrome PDF, then PNGs rasterised FROM the PDF (
pdftoppm -scale-to-x 1080 -scale-to-y 1350). The PNGs that get read are the pages that ship. Chrome rounds the page to 810x1013.04 pt, so-r 96gave 1080x1351; scaling to exact pixels fixed it. - Fit gate in the browser.
fit.jsruns under--dump-dom. It waits for the web fonts, shrinks only overflowing text, never below a per-role minimum (body 36px, title 52px on the 1080 canvas, which is about 12px and 17px on a phone), and measures contrast against the nearest opaque background.render.pybakes the fitted sizes into the HTML and re-verifies before printing, so the PDF never depends on JS timing. - Cover = blueprint grid; CTA =
--deep. These borrow the home intro's grid and the footer band's colour. Every other slide is paper. Field colour appears only inside the art frame ([[ART]]). The art reusesq-agent-trim.pngplus code-drawn tags, with no new Quiver calls ($0). - Draft marks. A slide's
benfield renders as the pack's dashed[BEN: ...]callout.--finalrefuses to render while any is open. - Hard checks are a script (
hardcheck.py, content-sdlc "Hard checks"): High-severity lexicon patterns (65 scannable; 15 structural entries left to the critic), weasel words, VOICE never-say phrases, hashtags, em-dash density, length limits per format, untraced first-person claims, open [BEN] tags and acronym expansion.
Dead ends / surprises
--user-data-dir=<tempdir>made headless Chrome 154 hang past 120s (a bisect showed the flag alone caused it). Removed.- The swipe arrow (U+2192) is not in Instrument Sans, so the PDF silently embedded Arial-BoldMT. The arrow is now an inline SVG, and
render.pynow fails any run whose PDF embeds a non-contract font. - The scratchpad is shared with a parallel agent building
/ideas. My first backup (scratchpad/SKILL.md.orig) was overwritten by its file. The real remix SKILL.md was untouched (10134 bytes, May 1). I re-backed it up toscratchpad/remix-builder/and did not touch the other agent's files.
Verification (how)
- Negative tests in the scratchpad: a 220-word slide gave
overflow/clip: content is 111px taller than the safe area, exit 1. An unbreakable 46-letter word gavetoo wide. A slide stuffed with "Let's dive in / ever-evolving / landscape / #agents / Ultimately / would help / drives / 3 em-dashes / untraced I" got 11 FAILs. The x format caught 5 sentences, a hashtag and "Ultimately,". - Sample render: 8 pages,
pdffontsshows only Source Serif 4 / Instrument Sans / JetBrains Mono, every PNG is 1080x1350, minimum contrast 5.3:1, hardcheck PASS with 4 open [BEN]. - Visual: all 8 PNGs read after every render (4 full reads). Pass log in [[remix]].
- Fresh-eyes critic (sw-critic, rubric passes 1-5 + fidelity): ITERATE. Fidelity and structural fixes were applied; the hook choice, "engineers" vs "advanced users" and the narrower answers went to the founder (see [[remix]]).
- Mode D (the original Sanity Check behaviour):
diffagainst the backup shows only heading levels (demoted one step, nested under Mode D) and the one "report via iMessage or Discord" line changed.
Not done / out of scope
- No git:
~/.claude/skillsis not a repo, so there is no PR. No pushes, no deploys, nothing posted. - [[CAPABILITIES]] is not updated (the brief forbids editing SOPs). It should list
/remixcarousel rendering when someone next regenerates it.