adam-j-ali/research-poster

<references_contract>

References are mandatory, not optional

This skill ships 6 reference files plus two worked examples in references/examples/. Read the matching reference BEFORE the step that needs it, never afterwards to check your work.

• references/data-integrity.md: reproduce the abstract from the source data (Phase 1)
• references/venues.md: find and date the venue's rules; institution brand kits (Phases 0 and 2)
• references/design-rules.md: readability floor, type scale, layout, house style (Phase 3)
• references/build-and-charts.md: native charts, the repair trigger, python-pptx pitfalls (Phase 4)
• references/rendering-qa.md: render, readability and overlap checks, QR codes (Phase 5)
• references/citations.md: verify, format, hyperlink (Phase 2) </references_contract>

This skill encodes a workflow that was hardened in practice. Read the matching reference BEFORE the step that needs it, not afterwards to check your work (by then the mistake is in the output); do not reinvent the parts that are already solved (chart-safety, rendering, typography scale). Scripts live in ${CLAUDE_SKILL_DIR}/scripts/; worked examples and deep guidance live in ${CLAUDE_SKILL_DIR}/references/. Rendering uses LibreOffice (cross-platform) or Microsoft PowerPoint on macOS; see references/rendering-qa.md.

<companion_skills> Everything this skill needs is bundled; the build, render, and check run on the scripts here alone. These optional skills add polish. Use any that are installed. If one is missing, carry on without it and mention the one-line install to the user once; never stop for it.

Skill	Adds	Install (in Claude Code)
Anthropic pptx + pdf (document-skills)	editing an existing poster, hard PDFs	/plugin marketplace add anthropics/skills, then /plugin install document-skills@anthropic-agent-skills
frontend-design	palette, type pairing, layout direction	/plugin install frontend-design@claude-plugins-official
K-Dense citation-management (MIT)	PubMed / DOI lookup, formatted references	copy skills/citation-management from github.com/K-Dense-AI/scientific-agent-skills into ~/.claude/skills/
K-Dense scientific-visualization (MIT)	journal-style matplotlib figures (image-chart route)	same repo, skills/scientific-visualization

Leave out K-Dense pptx-posters (an HTML/CSS poster route that competes for the same requests) and scientific-writing (manuscript prose that bans the bullets a poster needs). Citation checks need no extra skill: web search and fetch against PubMed / Crossref / doi.org do the job. </companion_skills>

<execution_context> Run interactively OR fully autonomously. Every gate below is an AskUserQuestion with a clearly marked recommended default, so an autonomous run just takes the defaults and states them.

• 🟢 Auto (no ask): read inputs, research, extract house style, build, render, self-QA and iterate, git-init + commit in the poster's own working dir, clean up intermediates.
• 🟡 Ask (with a recommended default): poster dimensions / venue spec, author + affiliation + logo confirmation, tone level, citation depth, chart engine (editable-native vs images), and whether the data is final.
• 🔴 Never: fabricate any number, result, or citation; alter the study's reported findings; overwrite a file the user has already finalized (always write a separate copy); ship a file that triggers PowerPoint's repair prompt; copy patient-level data out of the author's file or send it to any web service (read it in place, save aggregates only). </execution_context>

1. Collect inputs: the abstract PDF (required); an optional reference poster (.pptx) for house style; optional logo files; any updated data.
2. Reiterate what you understood, then batch the gates (one or two AskUserQuestion calls, each option marked with the recommended default). Recommended defaults in brackets:
   • Dimensions / venue: verify the venue's ePoster spec first [default]; else 16:9 single slide (48x27 in) which matches most on-screen ePoster viewers.
   • Authors, affiliations, and logos: confirm against the author, not the abstract's submission block (it is often not final). Multiple institutions means multiple logos [ask which].
   • Tone: moderate scholarly polish [default]; data and findings never change.
   • Citations: 4-6 verified references with clickable DOIs/URLs [default].
   • Chart engine: editable native charts [default]; polished images on request.
   • Is the data final, or is more coming? If ongoing, build so numbers are trivial to update (single data block) and keep the preliminary framing.
   • A single-slide poster or a talk deck? "PowerPoint for a conference" can mean either [poster].
   • When the author's data file is supplied and disagrees with the abstract: which numbers go on the poster [get the author's analysis file; else recompute and reproduce, see data-integrity.md].
   • Links for any QR codes (videos, full paper). Missing links become marked placeholders [default].
   • The upload deadline: ask for the acceptance email; a deadline may already be close.
3. Autonomous mode: take the defaults, state them, proceed.

1. Optionally use companion skills if the environment has them (document/pptx and pdf, dataviz, frontend design, citation management, scientific visualization). They are enhancements; the workflow below is self-contained via python-pptx, matplotlib, and a renderer.
2. Read the abstract completely: text via pdfplumber (python-docx for a .docx), a visual render, and an OCR/image check so no figure is missed. Capture the title, authors, affiliations, every number, and the n/N for each.
3. If the author supplied the underlying data, read its structure before any values, then reproduce every figure in the abstract from it in an analysis.py that asserts them and writes aggregates to data.json. Check SD vs SE, paired vs group means, each figure's own n, category labels against the original notes, and the first data row. See references/data-integrity.md.
4. If a reference poster is provided, extract its house style: slide size, palette (band + accent fills), fonts and sizes, section-bar structure, and the logo. Render it with bash ${CLAUDE_SKILL_DIR}/scripts/render.sh to see it. See references/design-rules.md.
5. Confirm python deps: pip install python-pptx matplotlib pdfplumber pypdfium2 pillow pypdf pymupdf.

1. Verify the venue's ePoster spec on the society's own pages for this year (aspect ratio, fonts, media, upload vendor, deadline, required disclosure wording) and date it. Confirm which organisation each source is about. See references/venues.md, which also covers taking the palette, typeface and logo from the institution's published brand kit.
2. Verify and fetch citations: real, PubMed/DOI-checked, with hyperlinks and access dates; include registry and grey-literature sources the author actually uses, not only journals. See references/citations.md. Never invent a citation.
3. Pull any background needed for the introduction. Consult current library docs when building.

1. Set house-style tokens, a column grid, clear visual hierarchy, poster-scale typography, breathable padding, and strong contrast. See references/design-rules.md.
2. Choose chart types honestly: doughnut/pie only for parts-of-a-whole; a bar chart for independent or non-mutually-exclusive proportions; big-number tiles for single figures. Label every chart with its n/N. See references/build-and-charts.md.
3. Define one editable data block that drives every chart and callout.

1. Adapt references/examples/build_poster_example.py (editable native charts) or references/examples/make_charts_example.py (matplotlib images) to the content and house style.
2. Build native editable charts with PowerPoint-safe styling ONLY. The single hard rule: never set a doughnut/pie data-label POSITION, that is the setting that makes PowerPoint repair-and-remove the chart. Per-slice fills, legends, percentage labels, number formats, and axis fonts are all safe. Put the doughnut percentage in the legend plus a big number in the hole. Full recipe in references/build-and-charts.md.
3. TDD-lite: assert every displayed percentage recomputes to the source number before saving.
4. Get authors, affiliations, and logos right per Phase 0.
5. Readability gate: set every size from one type-scale table with a floor (22 pt on a 48 in slide), set chart.font.size on every chart, and end the build by failing if any size in the saved file is below the floor (scripts/check_fonts.py). See references/design-rules.md.
6. Readiness gate: keep the unconfirmed items (author list, affiliation, disclosures, links) in a CONFIRMED checklist in the build. Until all are True, name the output ..._DRAFT.pptx, mark each gap visibly on the poster, and print what is still open.

1. Render with bash ${CLAUDE_SKILL_DIR}/scripts/render.sh <pptx> <out.pdf> <out.png> 1.7 (LibreOffice, or PowerPoint on macOS), then view the PNG.
2. For native charts, run python3 ${CLAUDE_SKILL_DIR}/scripts/check_charts.py <pptx>: it fails if any doughnut/pie carries a data-label position, the setting that makes PowerPoint delete the chart on open.
3. Run scripts/check_fonts.py <pptx> and scripts/check_overlap.py <rendered.pdf>, then downscale the render to 1920x1080 and view the densest regions as 1:1 crops. Decode every QR code from the render. See references/rendering-qa.md.
4. Fix against the references/rendering-qa.md checklist: no overflow or clipping, big enough type, breathable padding, high contrast, aligned columns, correct numbers. Iterate until it is genuinely good before showing the user.

1. Write a SEPARATE, clearly named copy; never overwrite a file the user has finalized.
2. Deliver the .pptx, a submission PDF, and a preview PNG.
3. git-init the poster's working dir if needed and commit with a clear message (baseline first when iterating on an existing poster, so the prior version is preserved).
4. Remove intermediate renders and temp files; keep the reproducible build scripts.

<success_criteria>

• Deliverable is an editable .pptx plus a matching PDF (verified page size) and preview PNG
• Charts are native and editable (default) and survive PowerPoint open with no repair prompt
• Every percentage recomputes to the source number; n/N shown; preliminary framing kept if ongoing
• Citations are all real and verified, with DOI/URL hyperlinks
• Authors, affiliations, and logos confirmed with the user
• Every figure the abstract states reproduces from the source data (when data was supplied)
• No text below the floor (check_fonts.py), no overlapping text (check_overlap.py), legible in 1:1 crops of the 1080p view; padding is breathable, contrast is strong, columns align
• Every QR code decodes from the render to exactly its link; placeholders keep the file a DRAFT
• House style matches the reference poster when one was provided
• A separate copy was written; the user's finalized files were not touched; work is git-tracked </success_criteria>

<error_handling>

• render.sh prints "RENDER FAILED" if no renderer is found: say so plainly and offer the two fixes (install LibreOffice, or use macOS with Microsoft PowerPoint). Do not surface raw AppleScript errors (-600 / -609 / -1712) to the user. The PowerPoint path already retries on a launch race.
• A PowerPoint render seems frozen: PowerPoint is usually waiting on a one-time macOS dialog (allow the terminal to control PowerPoint, or PowerPoint's "Grant File Access"). Ask the user to look at PowerPoint and approve it, then continue; do not kill and retry blind. The file-access grant is per folder, so render from a folder that has already worked rather than a fresh temporary one.
• render.sh exits 2 with "POWERPOINT IS OPEN": the PowerPoint path restarts PowerPoint without saving, so it refuses while the user has it open. Ask them to save and quit PowerPoint, then re-run. Never quit or kill PowerPoint yourself while their files are open.
• check_charts.py reports a REPAIR TRIGGER, or a native chart is missing from the render: a doughnut/pie data-label position was set somewhere; remove it (see references/build-and-charts.md).
• python dep missing: pip install python-pptx matplotlib pdfplumber pypdfium2 pillow pypdf pymupdf. </error_handling>