Academic Poster Builder

Paper Poster (HTML): measurement-gated poster generation

One HTML file styled for an exact print canvas (@page { size: W H }), rendered to PDF via Playwright print emulation. Iterate by measuring, not eyeballing — the screen preview lies; only print emulation at the correct viewport tells the truth. Core gate machinery is adapted from posterly (MIT, © 2026 Ruishuo Chen — see NOTICE.md and LICENSES/posterly-MIT.txt); ARIS adds style discipline gates, figure-provenance gates, the cross-model review loop, and the anti-patch-loop fix vocabulary.

Why this skill exists (the failure it prevents)

A predecessor pipeline produced a poster with 30+ colors, zero real paper figures, a screen-pixel canvas, and tiny formulas floating in oversized boxes, then spent 12+ review rounds making it worse — each round added a new badge color or bespoke SVG patch. The cure is structural, not exhortative:

1. Hard gates run before any aesthetic opinion (alignment, style, assets must PASS first — a reviewer never sees an unmeasured poster).
2. A closed fix vocabulary — visual-review fixes can only touch design tokens, whole catalogued components, content rebalance, assets, or canvas choice. New inline styles / new hex values / bespoke decorations are structurally forbidden.
3. Two-hue discipline as a machine check, not a style suggestion.
4. Real paper figures with provenance manifest, or the gate fails.

Mental model

paper (.tex / PDF) ──► content plan + claim→evidence audit (codex, fresh)
                              │
   figures extracted ─────────┤  FIGURE_MANIFEST.json (provenance, sha256)
   (real paper figures ONLY)  ▼
   template scaffold ──► fill ──► run_gates.py            ◄─── HARD, loop here
                                  preflight → style → asset → measure → polish
                              │ all hard gates PASS
                              ▼
                    Claude visual review (≤3 issues × ≤3 rounds, fix-vocabulary only)
                              │ score ≥ 9
                              ▼
                    codex final cross-model review (fresh thread, full HTML+PDF)
                              │ pass
                              ▼
                    verify-final → poster.pdf + GATE_REPORT.json

Constants

• SKILL_SCRIPTS = ${CLAUDE_SKILL_DIR}/scripts — all helpers are single-owner and ship inside this skill (Arch C). If the directory is missing the install is broken: abort and tell the user to re-install the skill (Policy A — the gates ARE the skill; never improvise replacements).
• REVIEWER_MODEL = gpt-6-astra, reasoning xhigh, fresh thread per review call (mcp__codex__codex, never codex-reply across review boundaries).
• CANVAS — from the venue's official spec, looked up live in Phase 0. Never assume. (Known anchor: ICLR 2026 main = 185×90 cm landscape per its official printing service; ICML/NeurIPS commonly 60×36 in landscape; workshop posters often 61×91 cm portrait. Specs change yearly — verify.)
• PALETTE — default = templates/tokens/generic.json (slate-blue #2D5F8B accent
  • gold #C9A24A highlight + neutrals) for all venues. Venue packs are opt-in via — venue-colors: true. Purple-dominant accents (hue 250–285) are banned unless the user passes — allow-purple: true.
• AUTO_PROCEED = false — wait for explicit confirmation at every 🚦 checkpoint.
• OUTPUT_DIR = poster_html/ in the working directory.

Workflow

Phase 0 — Resume, dependencies, venue spec

1. Resume: if poster_html/POSTER_STATE.json exists with status: in_progress (< 24 h), resume from the saved phase.
2. Dependencies (degradation chain, in order):
   • Playwright + bundled Chromium → if missing, python3 -m playwright install chromium → if install fails but system Chrome exists, scripts fall back to channel="chrome" → if all fail: you may produce the content plan and scaffold only, label everything "not print verified", and must NOT emit a final PDF.
   • pdfinfo missing → PyMuPDF reads PDF dimensions. At least one of pdftoppm / PyMuPDF must exist for PNG review renders.
   • MathJax: download tex-svg.js once into poster_html/assets/mathjax/ and reference it locally in the HTML. CDN is acceptable only for drafts; the measure gate hard-fails on unrendered MathJax either way.
3. Venue spec lookup (live): consult the venue's official poster-instructions page (search + fetch). Extract dimensions, orientation, font floor, logo policy, anonymity rules, file format. Record {spec, source_url, retrieved} into POSTER_STATE.json — specs change yearly; never reuse a cached spec silently.

🚦 Checkpoint: echo the venue spec table (canvas, orientation, source URL) and the chosen template. Wait.

Phase 0.5 — Design discovery (one AskUserQuestion batch)

Ask once, ≤4 questions: layout template (from templates/README.md), palette (default generic pack / venue pack / custom within constraints), logos + venue mark (paths or "none" — never fabricate; check the venue's logo policy), QR target (paper / code / project page / none — generate offline with qrencode or python-qrcode; never a remote QR-service URL). Persist answers in POSTER_STATE.json as design_decisions — re-read before any later "improvement" so deliberate choices are never reverted.

Phase 1 — Paper ingest, content plan, claim audit

1. Read the paper source (.tex ideal; PDF otherwise). Extract: title/authors/affils, the 3–5 headline numbers, core method (equations verbatim), main results (tables/figures and what they show), takeaways. Build poster_html/POSTER_CONTENT_PLAN.md — what goes in which column, word budget per card. Target density (excluding table cells, captions, author line, footer): standard poster 550–850 words; dense theory+empirical poster 750–1050 words, allowed only when ≥2 compact components are used (eqn-anatomy, flow-strip, derived-col, claim-pills, keybox--4). Warn yourself below 500 words on a 4-column landscape (it will read as sparse next to professionally dense posters) unless the template is hero/visual-first; warn above 1100 unless the user asked for dense mode. Bullets ≤ 8 words when possible — density comes from structure, not long prose. Prefer compact structure over prose: if the paper contains an explicit objective, algorithm, theorem mechanism, or baseline comparison, extract at least two of: (1) empirical objective / loss stack; (2) term-by-term equation anatomy; (3) a method-flow strip grounded in paper variables; (4) a derived-Δ column for method-vs-baseline rows; (5) a 4-up implementation/theory keybox; (6) a claim/evidence pill table for numeric-heavy posters. Do not invent an algorithm. If the paper has only an objective, label the component "objective flow" or "loss anatomy", never "algorithm".
2. Cross-model content audit (fresh codex thread, xhigh): give it the content plan path + paper source path(s) — paths only, no summaries — and ask for a claim→evidence table: | claim on poster | paper file:line | paper says (verbatim) | match? | with match ∈ {OK, NUMERIC-MISMATCH, OVERCLAIM, MISSING-PRECONDITION, NOT-IN-PAPER, SCOPE-NARROWED}. Save to poster_html/CLAIM_EVIDENCE.md.
3. Fix every non-OK row or record it as a user-acknowledged tradeoff.

🚦 Checkpoint: content plan + audit summary. Wait.

Phase 2 — Real paper figures (provenance-gated)

Source preference chain:

1. Paper source figures/ (vector SVG/PDF → convert to SVG via inkscape/pdf2svg if available, else rasterize ≥ 2× rendered px).
2. PDF-only: extract_pdf_figures.py contact-sheet + auto to list candidate regions → pick crops (🚦 human confirms crop choices) → crop at 300–450 DPI.
3. Last resort: user supplies explicit page,x0,y0,x1,y1 bboxes.

Then preprocess_figures.py --autocrop every asset. Every paper-derived image gets a FIGURE_MANIFEST.json entry (source hash, page, bbox, dpi, sha256, natural_px) and is embedded as <img data-source="paper" data-asset-id="...">.

Hard rule: ≥ 2 paper-derived visuals or the asset gate fails. Theory-only papers may waive the total-area rule (--waive-total-area) at a human checkpoint — never silently. Never draw bespoke decorative SVG "figures" as substitutes.

Figure-area bands (asset gate, fractions of body): total target 14–22 % (warn < 12 % / > 24 %, hard < 10 % / > 28 %); per ordinary figure target 4–8 % (warn

10 %, hard > 13 %); figure--duo combined 8–12 %. Hero templates pass --hero (centerpiece may take 30–40 %). The failure mode is symmetric: too small reads as decoration, too big crowds out content. Sibling figures that share axes or tell a before→after story belong in one figure--duo card, not two cards.

Phase 3 — Scaffold + tokens

cp templates/<chosen>.html poster_html/poster.html; retarget @page + .poster dims to the venue canvas (two edits, same values); apply the chosen token pack onto the :root DESIGN TOKENS block; fill content per the plan; embed manifest figures. Run preflight + style_check — both must PASS before any layout iteration. (A fresh scaffold is expected to fail measure — that gate judges a filled poster.)

Phase 4 — Layout hard loop

After every layout change:

python3 "$SKILL_SCRIPTS/run_gates.py" poster_html/poster.html \
    --tokens <pack.json> --manifest poster_html/FIGURE_MANIFEST.json \
    --report poster_html/GATE_REPORT.json

Canonical order: preflight → style → asset → measure → polish. Targets: column-bottom spread < 5 px (aim < 3), footer gap ∈ [30, 50] px, intercard gap ∈ [12, 50] px, canvas-fill ∈ [95, 101] %, poster bbox aligned to page within ±2 px. Fix guidance for each failure mode lives in the gate output and templates/COMPONENTS.md. Do not proceed while any hard gate fails. Do not let a reviewer see an unmeasured poster. Balance under-filled columns with content from the paper (Gate C), never with whitespace, space-between, or stretched cards.