Conference Poster

Conference Poster Builder

Turn a paper into a print-ready academic poster for any conference, with minimal effort and full control. The user supplies a paper (link or PDF) and a venue; the skill figures out the required poster size, builds the layout, and exports a PDF at the exact physical size. Output is one self-contained HTML file (editable live in the browser) plus the PDF.

What you require from the user (in this order)

1. The venue (FIRST, required): a link to the workshop / conference / paper page, or its

name. You then research the poster requirements for it (size, orientation, board/tape, print deadline).

1. The paper (required): a preprint link (arXiv / OpenReview / direct .pdf) or an

uploaded PDF of their paper.

1. The style (interview — see step 0C).
2. Optional: logos, exact author list, which figure is the centerpiece, QR target URL.

Workflow

Do these in order. Keep generated files in a working dir like ./poster// (create it; never write into the skill folder). Scripts live in this skill's scripts/. Use AskUserQuestion for the interview; ask in this sequence, don't interrogate, and skip anything already given.

0A. Venue → research the poster spec (ask first)

Ask for the workshop / conference / paper link or name. Resolve its poster spec from reference/poster-specs.md: presets table first; else web-search the official instructions, extract orientation + max W×H (convert to inches) + board/tape/deadline rules. Confirm the final WIDTH × HEIGHT + orientation with the user (a wrong size = a wasted print). Fallback: A0 portrait (33.1×46.8 in), flagged.

0B. Paper

Ask for the preprint link or a PDF. (Fetched in step 1.)

0C. Style interview (formal vs informal, then branch)

Ask whether they want it formal or informal / fun — let them describe in their own words what they want; if they're unsure, propose a style (offer 2–3 concrete ideas, incl. a relevant association with their topic). Once they agree, branch:

• FORMAL. Ask them to describe the design approximately — fonts and colors. Build a

clean academic poster: their palette (or a sober preset: Indigo / Ocean-ICML blue / Slate mono / Crimson), the requested fonts, no .decor, no emoji, no slang, logos only, lots of whitespace ("Clean academic mode" in reference/styles.md).

• INFORMAL / fun. Ask if they have a fal.ai API key (offer it — it gives by far the

best result; get it via AskUserQuestion "Other" → paste, or FAL_KEY; never hard-code/log it).

• No key → web-preset flow (like the original Floppa build). Pull a real, trendy palette

(scripts/extract_palette.py / curated-Pantone, cite it) and real themed art from the web (scripts/fetch_image.py), and place the images only where there is no text — margins / whitespace / beside figures, composed HERO-first (decoration-forward layout in reference/styles.md). Never emoji-only from memory.

• Key given → paint-over flow (best). (a) Build a clean base poster first — nice

colors, no decor/emoji, roomy empty margins (.poster padding ~1.4in) — and rasterize it to base.png. (b) Hand it to Nano Banana Pro to paint cohesive themed decoration in the user's style: scripts/decorate_poster_fal.py base.png decorated.png --theme "" --resolution 4K. (c) ⚠️ The image model corrupts text/numbers — overlay the pixel-exact content back: scripts/composite_content.py index.html decorated.png poster_final.png (transparent re-render, decoration stays integrated, data stays exact). (d) Wrap to PDF: scripts/img_to_pdf.py poster_final.png poster.pdf .

0D. Logos (optional)

Ask which org/lab logos to include. The user can name them; fetch each with python3 scripts/fetch_logo.py ./poster//logo-.png (web-search the official domain / a Wikimedia SVG for a clean result). Drop them into the header .logos block.

1. Get the PDF + real metadata

• PDF. If given a URL or arXiv id: `python3 scripts/fetch_pdf.py

./poster//paper.pdf` (resolves arXiv/OpenReview to the real PDF; stdlib-only). If given a local path: use it directly. If a download fails (paywall/landing page), ask the user to drop the PDF in.

• Authors (important). If the paper is on OpenReview, the PDF is often anonymized

but the real author list is public — fetch it: python3 scripts/fetch_openreview_meta.py returns title, full author list, and venue. Use these authors verbatim (don't trust an "Anonymous Authors" PDF). For arXiv, the PDF/abs page authors are reliable.

2. Extract content + figures from the PDF

Pull text and figures. Helpers are PyMuPDF-based and need no poppler:

• Read the PDF with the Read tool to get title, authors, affiliations, abstract,

contributions, method, key results, references.

• Figures shipped as separate files (often vector .pdf): rasterize each, auto-trimmed:

python3 scripts/fig_to_png.py fig.pdf ./poster//centerpiece.png --dpi 300 (works on .pdf pages and .png/.jpg; trims whitespace).

• Only the paper PDF: python3 scripts/extract_figures.py paper.pdf ./poster//_figs

dumps embedded rasters + renders every page with a manifest; pick a centerpiece, then clean it with fig_to_png.py.

• Legacy fallback if poppler is present: pdfimages -all / pdftoppm -png -r 200.
• If extraction fails entirely, proceed with placeholders and ask for figure PNGs.

Capture specifically:

• Title, authors, affiliations, venue line.
• The single headline finding in plain language → the giant central takeaway. A

complete sentence a non-expert grasps in 5 seconds. This carries ~90% of the poster — spend real effort here.

• 3–5 contributions / key results as short bullets (with the actual numbers).
• A 1–2 sentence method summary.
• The centerpiece figure + 1–3 supporting figures.
• 3–6 key references (short form).

3. Build the poster HTML at the right size

Copy assets/template.html into the working dir as index.html, then:

1. Set the physical size from step 0 in BOTH places (they must match): the :root

--w / --h variables AND the @page { size: } rule. Use real units (in/cm).

1. For landscape posters, widen .support to 3 columns (noted inline in the template).
2. Fill the marked `` sections with the paper content.
3. Apply the style from step 0C: set the :root palette + fonts; for formal and for

the paint-over (informal+key) flows build this as the clean poster (no .decor, no emoji; paint-over also uses roomy ~1.4in margins). For informal without a key, also place the web-sourced art in the .decor layer (HERO-first, only where there's no text).

Design rules to honor (Morrison "Better Poster" + ML-conference norms — see reference/design-rules.md):

• One giant plain-language takeaway dominates the center (the hero).
• Minimal text. ~40–50% of area = visuals/whitespace. Cut ruthlessly. The poster is

the tools for a 5–10 min pitch, not a copy of the paper.

• Type sizes scale with poster size; the template's CSS vars are tuned for ~24×36 —

bump them up for A0/larger. Body must be readable from ~4–6 ft.

• QR ≥ ~4 in, near the takeaway, labeled "Scan for the paper". Sans-serif, one accent

color, generous margins.

Generate the QR with python3 scripts/make_qr.py ./poster//qr.png (auto-installs segno). Embed it (the template references qr.png).

4. Render the PDF at the exact size

bash scripts/export_pdf.sh ./poster//index.html ./poster//poster.pdf (pass the target width/height in inches so it verifies the output). Uses headless Chrome honoring the CSS @page size, then checks it's a single page at the expected dimensions. If size/page count is wrong, fix @page/:root to match and re-run.

For the paint-over (informal + fal.ai key) flow, this export is the clean base: rasterize it to base.png, then run the decorate → composite → img_to_pdf chain from step 0C to produce the final decorated poster.pdf.

4b. Design review (composes with the Anthropic "Design" plugin)

Before handoff, run a design + accessibility pass on the rendered poster — see reference/design-review.md.

• Render a preview.png (e.g. `python3 -c "import fitz; fitz.open('poster.pdf')[0]

.get_pixmap(matrix=fitz.Matrix(110/72,110/72)).save('preview.png')"`) and look at it.

• Always run python3 scripts/contrast_check.py ./poster//index.html — it audits

the :root colors + .takeaway gradient stops against WCAG (a vibrant band gradient often fails white text at its lightest stop; darken that stop). Fix failures, re-export, re-check.

• If the Design plugin is installed, invoke its design-critique and

accessibility-review skills on the preview/HTML and apply the high-severity findings (hierarchy, consistency, contrast, legibility). If not, use the distilled checklist in reference/design-review.md. Then re-export.

5. Hand off

Report to the user:

• Paths to index.html (editable) and poster.pdf (print-ready), and the final size.
• How to edit: open index.html in Chrome; text is editable inline; swap images by

replacing files; tweak accent color / sizes via the CSS variables; re-run step 4.

• Print/logistics: the size + orientation, the print deadline and mounting rules you

found in step 0 (tape vs pins, board location), and recommended 100–300 DPI at full size. For ICML workshop specifically, see reference/icml-2026-spec.md.

• ⚠️ Proofread before printing: AI can misrender equations, numbers, citations, and

the poster size. Tell the user to verify every metric, figure label, author name, and the physical dimensions against the official venue page.

Files

• assets/template.html — self-contained billboard poster template (inline CSS),

parametric size (set :root --w/--h + @page); portrait or landscape.

• scripts/fetch_pdf.py — resolve a preprint URL / arXiv id to a local PDF (stdlib only).
• scripts/fetch_openreview_meta.py — title + full author list + venue from an OpenReview

submission (the real authors, even when the PDF is anonymized). Stdlib only.

• scripts/fetch_logo.py — download an org logo as PNG by name / domain / image URL

(Clearbit + favicon fallback). For a name, web-search the real domain first.

• scripts/fetch_image.py — download themed decoration art (transparent PNG/SVG) by direct

URL(s) for "described" styles; preserves transparency, resizes sanely.

• scripts/extract_palette.py — pull a real, on-trend color palette (hex + role

suggestions) from a reference image / the hero asset, for the style-discovery flow.

• scripts/gen_asset_fal.py — OPTIONAL: generate bespoke transparent-PNG themed art via

fal.ai Nano Banana Pro (Gemini 3 Pro Image) when the user supplies a FAL_KEY; auto-cuts the rendered checkerboard with rembg; exits non-zero (fall back) if no key.

• scripts/decorate_poster_fal.py — OPTIONAL (preferred themed flow): hand a clean rendered

poster PNG to Nano Banana Pro (edit) to paint cohesive themed decoration into the margins.

• scripts/composite_content.py — re-render the clean HTML with a transparent background and

overlay it on the decorated image, so data stays pixel-exact AND the decoration stays integrated (not boxed). Mandatory after decorate (the edit model corrupts text/numbers).

• scripts/img_to_pdf.py — wrap a full-bleed poster image into a single PDF page at exact size.
• scripts/extract_figures.py — dump embedded images + render pages from a paper PDF

(PyMuPDF, no poppler), with a manifest, to pick a centerpiece.

• scripts/fig_to_png.py — rasterize one figure (vector PDF page or image) to a

whitespace-trimmed, high-DPI PNG ready for the poster.

• scripts/make_qr.py — QR PNG generator (auto-installs segno).
• scripts/contrast_check.py — WCAG 2.1 contrast audit of the poster's palette (parses

:root + the .takeaway gradient); also checks a single fg/bg pair.

• scripts/export_pdf.sh — headless-Chrome HTML→PDF at the poster's exact size; verifies

single page + dimensions (PyMuPDF fallback when poppler is absent).

• reference/poster-specs.md — how to research any venue's spec + a presets table.
• reference/styles.md — named color/style presets + the research-driven "described styles".
• reference/design-review.md — design + accessibility review pass; composes with the

Anthropic "Design" plugin (design-critique / accessibility-review) or runs a distilled print-poster checklist standalone.

• reference/icml-2026-spec.md — detailed ICML 2026 workshop spec + logistics (one preset).
• reference/design-rules.md — billboard / Better-Poster design checklist.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: zj-karina
• Source: zj-karina/my-little-poster
• License: MIT

Install and usage instructions live in the source repository linked above.