How to use the Academic Poster Builder (HTML) agent skill

What the skill makes the agent do

The skill turns a paper (a .tex source or a PDF) into a conference poster for venues such as ICML, NeurIPS, ICLR or CVPR. The poster is a single HTML file styled for the venue's canvas size using an @page size rule, and it is converted to PDF through Playwright print emulation.

Before a design is shown, the skill runs hard gates covering preflight, style, assets and measurement, and a polish step. The agent iterates on failures instead of relying on how the layout looks on screen, since the skill's documentation notes that the screen preview can differ from print output.

Figures are meant to come from the real paper. The skill extracts them from the PDF and records a FIGURE_MANIFEST.json with provenance and sha256 hashes, and the asset gate fails if that record is missing.

• Content plan written to poster_html/POSTER_CONTENT_PLAN.md before layout
• Default colour scheme: slate blue and gold, with purple-dominant accents blocked unless you opt in
• Output is poster_html/ containing the PDF and GATE_REPORT.json after verification

Install the skill (ZIP download or Claude apps)

Download the skill ZIP from the directory page and unzip it. For a personal install, place the skill folder in ~/.claude/skills/. For a single project, place it in that project's .claude/skills/ folder.

If you use the Claude apps instead, open Settings, then Capabilities, then Skills, and upload the skill there.

• Personal install: ~/.claude/skills/paper-poster-html
• Project install: .claude/skills/paper-poster-html inside your project
• Apps: Settings → Capabilities → Skills → upload

Install the tools the gates depend on

The gates need Python 3, Playwright with a bundled Chromium, and a PDF dimension reader. If Chromium is missing, the skill's documented command is python3 -m playwright install chromium. If that fails but system Chrome is installed, the scripts fall back to system Chrome.

pdfinfo or PyMuPDF must be available to read PDF dimensions, and at least one of pdftoppm or PyMuPDF is needed for PNG review renders. Optional extras are qrencode or the Python qrcode package for QR codes, and a locally downloaded MathJax tex-svg.js for formulas.

• Required: Python 3, Playwright with Chromium (or system Chrome)
• Required for PDF dimensions: pdfinfo or PyMuPDF
• Optional: qrencode or python qrcode for offline QR codes
• Optional: MathJax tex-svg.js saved into poster_html/assets/mathjax/

Trigger the skill with a natural request

The skill activates when you use phrases such as poster, conference poster, paper poster, or the Chinese phrase 做海报, or when you ask the agent to design or redo a research poster. Give it the paper location and, if you have them, the venue and canvas details.

The skill's argument hint shows the optional settings it accepts: the paper directory or PDF, the venue, the canvas size and orientation, and whether to use venue colours.

• Make a conference poster from ./paper/main.tex for ICLR
• Build a paper poster from this PDF, 185x90cm landscape, venue colours on
• Redo my research poster in HTML and export it as a PDF
• Argument hint format: [paper-dir-or-pdf] [— venue: ICLR, canvas: 185x90cm landscape, venue-colors: true]

What happens during a run

The agent first looks up the venue's official poster instructions for the canvas size, orientation, font minimum and file format, and records the source. It then asks a short batch of design questions covering layout template, palette, logos and QR target. It does not assume a canvas size.

After the content plan and figure extraction, it fills a template and runs the gates. Only after the hard gates pass does a visual review happen, limited to a few issues per round and to a fixed set of allowed fixes. Checkpoints wait for your confirmation before the agent proceeds.

• Venue spec is looked up live and saved with its source URL
• Checkpoints pause for your confirmation
• Visual review is limited to up to 3 issues per round and 3 rounds
• Final output is verified before poster.pdf is emitted

Safety notes: scripts and network access

The skill runs scripts and uses the network. Its helpers ship inside the skill folder, and the skill says it will not replace them with improvised versions. If the scripts folder is missing, the skill stops and asks you to reinstall it.

Network use includes looking up the venue's poster page, downloading MathJax tex-svg.js once if you choose to, and an optional cross-model review through an MCP reviewer. QR codes are generated offline, not through a remote service. Review the scripts folder and the allowed tools in SKILL.md before you install if you have restrictions on what runs on your machine.

• Scripts: under scripts/ (including scripts/_posterly/ and asset_check.py, extract_pdf_figures.py)
• Network: venue spec lookup, optional MathJax download, optional reviewer over MCP
• Licence: MIT; posterly-derived gate code is credited in NOTICE.md and LICENSES/posterly-MIT.txt

Tips for better results

Give the agent the .tex source when you have it, because the skill prefers it over PDF for extracting equations and text. Check the venue's own poster rules yourself as well, since the skill's canvas defaults are only anchors and specs change each year.

Keep the word count in the target range. The skill suggests 550 to 850 words for a standard poster and warns below 500 words on a four-column landscape layout. Avoid asking for custom artwork, because the skill's fix list restricts bespoke decoration by design.

• Provide the .tex source when available
• Confirm the canvas size against the venue's official page
• Aim for 550 to 850 words on a standard poster
• Ask for logos and QR targets explicitly, and the agent will not invent them