D3 Poster (Conference / Research)

<skill-dir> in this skill means the skill's own base directory, which Claude Code shows when the skill loads (installed as a plugin: ${CLAUDE_PLUGIN_ROOT}/skills/d3-poster; copied by hand: ~/.claude/skills/d3-poster). Substitute the real path in every command. <wording-dir> means <skill-dir>/../d3-scientific-wording, the shared wording skill that every D3 plugin ships.

Create professional A0-portrait research posters with D3 (Data Driven Decisions) and University of Würzburg branding in the university's 2026 corporate design. Builds on the baposter class for a grid layout, with a thin D3 reskin (d3-poster.sty) that draws the frame of the university poster template (navy header box with the white university logo, white title and authors, and the blue seal; grey section bars with navy text; a footer with a navy rule, a logo row with the D3 logo first, and a references slot), sets the D3 palette (university navy, D3 orange accent), and ports the \highlight / \keypoint / infobox / card macros from d3-abstract. Compiles with XeLaTeX or LuaLaTeX for the house font Inclusive Sans UW; falls back to pdfLaTeX with a Helvetica clone.

The seal in the header is an interim choice: the university's CD booklet replaces it with a faculty ornament in faculty and chair communication, and that faculty package is not released yet. Swap uw-seal-blue.pdf once it is.

Reference files (loaded on demand)

The skill body below is the workflow; long reference material lives in references/. Load a reference file when its topic comes up:

File	Load when
`references/extraction.md`	About to extract content from a paper (any tier). Has the tier-by-tier extraction commands, PDF figure extraction (vector-first), low-res raster protocol, on-disk format priority, architecture-figure interview.
`references/tikz-recipes.md`	About to draw a figure from scratch in TikZ. Has 4 verified-compiling recipes (block diagram, pipeline, layered architecture, comparison) plus general TikZ rules.
`references/error-handling.md`	Compile errors or visible rendering bugs. Has the 8 known issues (missing class, missing fonts, siunitx `\d` conflict, header gradient instead of the solid fill, Calibri error, silent clipping, vanished Greek letters or symbols, title block set smaller).
`references/dual-agent-review.md`	Review depth is `standard` or `full` (see Review). Has the reviewer prompts, the iteration loop for `full` (capped at 4 rounds), score-to-README protocol.
`references/macros-and-styling.md`	Writing the poster body and need to confirm macro names, color values, font sizes, table conventions.

Template files and assets

<skill-dir>/
├── SKILL.md                    # this file
├── baposter.cls                # class file (copied to each project's drafts/)
├── example-poster.tex          # reference example -- DO NOT copy into user projects
├── references/                 # progressive-disclosure references (see table above)
└── assets/                     # this entire folder IS copied into user projects
    ├── d3-poster.sty
    ├── logos/                  # uw-logo-white.pdf, uw-logo-navy.pdf, uw-seal-blue.pdf, logo-d3.pdf, DataDrivenDecisions.png
    └── fonts/                  # InclusiveSansUW-{Regular,Italic,Medium,SemiBold,Bold}.ttf, OFL-InclusiveSansUW.txt

Two placement rules:

baposter.cls sits at the SKILL ROOT, not under assets/. \documentclass does not reliably accept path prefixes, so the class file must be a sibling of the .tex file at compile time.
example-poster.tex sits at the SKILL ROOT, NOT inside assets/. Never copy it into a user project. It's a reference you read when you need to remember the patterns. If you copy assets/ wholesale, you don't pick up example-poster.tex — exactly right, because shipping it alongside the user's poster.tex would confuse Overleaf's "Main document" selector and leave a stale example-poster.pdf. The user's project should contain ONE .tex whose stem matches its .pdf.

Project setup (save location + folder structure)

Save location priority

Posters are usually tied to a paper. Where you save determines whether collaborators can compile on Overleaf when local LaTeX isn't available. Decide the save location BEFORE creating any files:

Paper submodule detected → <submodule>/poster/. If .gitmodules lists a submodule that contains a paper (path like paper/, manuscript/, overleaf/, or any submodule whose contents include a \title{} + \begin{abstract} .tex), save inside that submodule. Lab convention: paper submodules live on GitHub and the corresponding Overleaf project syncs via Overleaf's GitHub Sync — the remote will be github.com/..., not git.overleaf.com.
No submodule but a paper folder exists → <paper-folder>/poster/. Still place under the paper folder.
Fallback → ./poster/ at the current working directory root.

Detect with:

test -f .gitmodules && git submodule status
for sm in $(git config --file .gitmodules --get-regexp path 2>/dev/null | awk '{print $2}'); do
  find "$sm" -maxdepth 3 -name "*.tex" -exec grep -l "\\\\begin{abstract}" {} \; 2>/dev/null
done

Print and confirm with the user before creating files:

[HH:MM:SS] SAVE LOCATION: <chosen-path>/poster/
  Reason: [paper submodule at paper/ (GitHub-hosted, Overleaf-synced) | paper/ folder exists | repo root fallback]
  Reachable on Overleaf? [yes -- push submodule to GitHub, Overleaf will sync | no -- local compile only]

Folder structure

<save-target>/poster/
├── progress.md
├── SUMMARY.md
├── drafts/
│   ├── baposter.cls            # copied from skill (sibling of poster.tex)
│   ├── assets/                 # copied from skill (d3-poster.sty, logos, fonts)
│   ├── figures/                # optional: poster-specific figures
│   ├── poster.tex              # the one working file (revisioned via git)
│   ├── poster.pdf              # always has the matching stem
│   └── revision_notes.md
└── final/
    ├── baposter.cls            # duplicated for Overleaf-standalone upload
    ├── assets/                 # duplicated for Overleaf-standalone upload
    ├── poster.tex              # same filename as in drafts/ -- stem invariant
    ├── poster.pdf
    └── README.md               # date, title, venue, review scores

Setup commands:

POSTER_DIR=<resolved-path>/poster
mkdir -p "$POSTER_DIR/drafts" "$POSTER_DIR/final"
cp    <skill-dir>/baposter.cls "$POSTER_DIR/drafts/baposter.cls"
cp -r <skill-dir>/assets       "$POSTER_DIR/drafts/assets"
# Do NOT copy example-poster.tex -- it stays at the skill root.

Naming rule (critical for Overleaf): the working file is always poster.tex, producing poster.pdf. Stem matches. Versioning uses git, not filename prefixes — no v1_poster.tex, no YYYYMMDD_title.tex. A previous version of this skill used such prefixes; in practice it confused Overleaf uploads (the user would set v2_poster.tex as main doc, but the cached v1_poster.pdf next to it made it unclear which was the real output).

Why duplicate baposter.cls + assets/ in both drafts/ and final/? ~1 MB per copy, but it guarantees each folder is independently uploadable to Overleaf — a collaborator opening final/ doesn't need anything from drafts/.

Document skeleton

Every poster MUST start with this structure:

%! TEX program = xelatex
\documentclass[a0paper]{baposter}

\usepackage{relsize}
\usepackage{url}
\usepackage{multicol}
\usepackage{amsmath,amssymb}
\usepackage{natbib}
\usepackage{graphicx}

\usepackage{assets/d3-poster}   % D3 reskin: colors, fonts, macros

\DeclareMathOperator*{\argmin}{arg\,min}
\renewcommand{\vec}[1]{\boldsymbol{#1}}   % Optional: bold math vectors

\begin{document}

\begin{poster}{
  columns=2,                     % DEFAULT 2. Use 3 ONLY when content truly justifies it.
  grid=false,
  eyecatcher=false,
  % University frame from d3-poster.sty: 25 mm margins, navy header box,
  % footer with logo row and references. These two lengths fit the template.
  headerheight=\dthreeheaderheight,
  colspacing=\dthreecolspacing,
  background=plain,
  bgColorOne=white,
  bgColorTwo=white,
  % Section bars: #ECECEC with navy SemiBold text. Per box, the key
  % `emphasis` gives the cornflower bar with white text.
  borderColor=d3-headergray,
  headerColorOne=d3-headergray,
  headerColorTwo=d3-headergray,
  headerFontColor=d3-navy,
  headerfont=\dthreesectionfont,
  headershape=rectangle,
  headershade=plain,             % solid headerColorOne fill; baposter's default `shadeLR` is a
                                 % gradient. See references/error-handling.md #4.
  headerborder=none,
  boxColorOne=white,
  boxshade=plain,
  boxpadding=0.6em,
  textborder=none
}
%%% Eye Catcher (empty; the university logo takes this place) %%%
{}
%%% Title -- plain text; set white Bold in the navy header box. No subtitle by default.
{Your Poster Title}
%%% Authors -- one line, white, below the title.
{Author Name, Data Driven Decisions Group\enspace|\enspace [email protected]}
%%% Footer logo row: D3 logo across the first two cells
{\dthreelogos}

%%% Motivation (column 0, row 0)
\headerbox{Motivation}{name=Motivation,column=0,row=0}{
  Lead paragraph: why this work matters, what's new, headline metric.
  Use \highlight{} sparingly on key terms.
}

%%% Problem (column 0, fills to bottom)
\headerbox{Problem}{name=Problem,column=0,below=Motivation,above=bottom}{
  Formal problem setup. Domain-appropriate math.
  \keypoint{One-sentence headline takeaway.}
}

%%% Approach (column 1, top)
\headerbox{Approach}{name=Approach,column=1,row=0}{
  Methodology. End with the method-overview figure (no caption needed -- describe in body text).
}

%%% Results (column 1, fills to bottom)
\headerbox{Empirical Results}{name=Results,column=1,below=Approach,above=bottom}{
  Infoboxes with metrics + a booktabs comparison table.
}

%%% Literature -- footer references slot (3 references, no separate box)
\dthreereferences{
  Author1 (Year, \textit{Venue}); Author2 (Year, \textit{Venue}); Author3 (Year, \textit{Venue}).
}

\end{poster}
\end{document}

CRITICAL: Use \usepackage{assets/d3-poster} (with assets/ prefix) — the wrapper expects to be loaded from a project root that has assets/d3-poster.sty and a sibling baposter.cls.

What d3-poster.sty draws (from the university's A0 poster template; geometry scales with the paper, so A1 and other sizes keep the proportions):

25 mm margins. Navy header box across the full width, 119 mm high, with the white university logo top left, the title (white Bold, 50 pt on A0, as in the template) and the authors (white, 34 pt) in the template's text field, and the blue seal flush with the box's right and bottom edges.
Section bars about 24 mm high: #ECECEC with navy SemiBold text. Add emphasis to a box's keys for the template's second style, a cornflower bar with white text.
Footer below the columns: a navy rule; on the left a logo row of 72 × 39 mm cells filled from the 5th argument of \begin{poster} (\dthreelogos = the D3 logo across the first two cells, about 30 mm high; \dthreelogocell{file.pdf} adds a partner logo, at most three more); on the right the References slot (\dthreereferences{...}).
Earlier posters still compile: their option block keeps the old look of the bars, the header and footer are drawn anyway, and \dthreelogos now fills the footer logo row instead of the header.

Box positioning, columns, title

Column count — default 2, 3 only if needed

Use columns=2 by default. Use columns=3 ONLY when:

The paper has 5+ truly distinct content sections that can't be combined (rare)
The user explicitly requests it for a venue that expects it

In every layout, put the Literature into the footer's References slot with \dthreereferences{...} (a short list, e.g., "Author1 (Year, Venue); Author2 (Year, Venue); ..."), not into a box. The university template reserves the right half of the footer for it, and it avoids the cramped Literature box that 3-column tends to produce at A0 portrait. 3-5 short references fit; a longer block is set smaller and the log says so.

Default 2-column grid

+-----------------------------------------------+
| [uni logo]  Title, authors (white)     [seal] |  navy header box
+-----------------------+-----------------------+
| Motivation            | Approach              |
| (col 0, top)          | (col 1, top)          |
+-----------------------+-----------------------+
| Problem               | Empirical Results     |
| (col 0, fills to      | (col 1, fills to      |
|  bottom)              |  bottom)              |
+-----------------------+-----------------------+
| -------------------- navy rule -------------- |
| [D3 logo] [partner]   | References            |  footer
+-----------------------+-----------------------+

3-column variant (rare)

+---------------+---------------+---------------+
| Motivation    | Approach      | Empirical     |
| (col 0)       | (col 1,       | Results       |
+---------------+  full height) | (col 2,       |
| Problem       |               |  full height) |
| (col 0)       |               |               |
+---------------+---------------+---------------+
| footer: logo row (D3 first)   | References    |
+---------------+---------------+---------------+

`\headerbox` keys

Key	Purpose	Example
`name=X`	Unique identifier for cross-refs	`name=Motivation`
`column=N`	0-based column index	`column=0`
`row=N`	0-based row from top	`row=0`
`below=Y`	Position directly below box Y	`below=Motivation`
`above=bottom`	Stretch to page bottom	`above=bottom`
`height=1`	Fill full column height	rarely needed in 2-col
`span=N`	Span N columns	`span=2` (full-width banner)

Title — in the navy header box, no subtitle by default

Pass the title as plain text (see example-poster.tex): d3-poster.sty sets it white, Bold and ragged-right in the template's text field between the university logo and the seal (425 mm wide on A0, 50 pt). Keep titles short enough to fit on 1-2 lines. A title block that does not fit the header box is set smaller and the log shows d3-poster Warning: The title block was set at ...; shorten the title instead of accepting that.

Subtitle: do NOT add one by default. Single-line punchy title is the default. Add a subtitle only if the user asks; see references/macros-and-styling.md for the syntax (\dthreesubtitle{...}).

Paper / repo scanning workflow

Posters almost always derive from an existing paper, but the input setup varies a lot. Some users have a full repo with paper submodule + figures; others have only a PDF a coauthor sent them; some have just notes. Detect the tier before deciding how to extract.

Step 1: Detect the input tier

Probe the working directory and stop at the first match:

Tier	What you find	How to detect
A. Full repo + paper submodule	Outer repo with `.gitmodules` pointing at a paper repo on GitHub, alongside `figures/`, `results/`, `code/`, etc. The canonical case.	submodule contains an abstract-bearing .tex AND sibling `figures/` / `results/` / `code/` exist
B. Paper repo only	A single repo or submodule containing .tex, figures, references — nothing outside. Common for Overleaf-only projects.	submodule (or cwd) has abstract-bearing .tex + paper-relative `figures/`, but no sibling code/results dirs
C. Single paper file	Just a `.pdf` or `.docx` in cwd, no source. Common when a coauthor emails the compiled paper.	a paper file but no .tex, no submodule
D. Notes / topic only	No file. User describes the work inline.	none of the above

Print [HH:MM:SS] INPUT TIER: <X> -- <one-line description> so the user can confirm or override.

Step 2: Extract content

Load references/extraction.md for the tier-specific commands. Summary:

Tier A/B: parse the .tex source — the extraction-mapping table maps \section{...} → poster boxes, \begin{table} → booktabs tables, \includegraphics{...} → figure inventory, references.bib → footer References slot.
Tier C (.pdf): dump PDF text via uv run --with pymupdf; for figures, prefer the PDF-page-crop approach (vector-preserving) before falling back to raster embedding extraction or full-page high-DPI render.
Tier C (.docx): dump text and images via uv run --with python-docx.
Tier D (no file): skip extraction; interview the user with the structured prompt template in references/extraction.md.

Step 3: Scan for figures, results, code

Asset type	Default search paths	Format priority
Figures	`figures/`, `figs/`, `plots/`, `images/`, paper-folder/figures/	PDF > EPS > SVG > PNG > JPG — prefer vector. Dedup by stem (see `references/extraction.md`).
Generated results	`results/`, `output/`, `experiments/`, `runs/`	Same: PDF preferred
Code / methodology	`code/`, `src/`, `notebooks/`, `scripts/`	n/a
Paper-relative	Anything inside the paper subfolder	Same as Figures

Quick survey:

for d in figures figs plots images results output experiments runs code src notebooks scripts; do
  [ -d "$d" ] && echo "Found: $d/ ($(ls $d | wc -l | tr -d ' ') items)" && ls "$d" | head -5
done

If none exist, broaden the scan (find . -type f $ -name "*.pdf" -o -name "*.png" $ -not -path "./.git/*") and ask the user which assets are poster-relevant.

Why vector first: A0 prints at ~84 cm wide; even a small figure on a poster is 25+ cm. Raster at paper-column resolution (~800 px) becomes 30 px/cm at A0 — visibly pixelated. Vector renders at print resolution regardless of physical size.

Step 4: Print a content brief

After Steps 1-3, synthesize a brief that stays in conversation context:

Paper found: paper/manuscript.tex
- Title: <extracted title>
- Abstract -> Motivation lead: <2-3 sentence condensation>
- Sections: Intro -> Motivation; Setup -> Problem; Method -> Approach; Results -> Empirical Results
- Figures inventoried: 6 in figures/ (top candidates: regret-vs-n.pdf, method-overview.pdf)
- Tables inventoried: 2 in paper/ (main_comparison.tex looks like the headline)
- Bib hits: 47 references; will narrow to 3-5 for the footer References slot

Want to use this as the seed, or override anything?

Get user approval before mapping content into boxes.

Creating figures: decision tree

When you need a figure in a box, choose:

A vector figure already exists on disk (PDF/EPS/SVG, same stem dedup): use it directly. \includegraphics[width=\linewidth]{...}.
A raster figure exists but at sufficient resolution (≥2200 px at column width): use it; warn if marginal.
A raster figure exists but at low resolution (<1500 px at column width): offer to redraw. See references/extraction.md for the threshold derivation.
An architecture-style figure was detected (filename or caption contains "architecture", "pipeline", "overview", "framework", ...): interview the user before placing it (use-as-is / simplify / replace / skip). See references/extraction.md §8.
No usable source figure: draw in TikZ. Load references/tikz-recipes.md for 4 verified recipes (block diagram, pipeline, layered architecture, comparison schematic) + general TikZ rules.

TikZ output is vector by default — the cleanest path when redrawing.

Workflow protocol

Phase 1: Plan

Analyze the request. New poster or editing an existing one? Identify the paper source (run Step 1 of paper scanning). Print: [HH:MM:SS] PLANNING: [new|editing].
Optional knowledge enrichment. Offer only when the user's request is high-level, no paper is linked, or the user mentions "literature", "recall", "past work". Skip if a paper was scanned or the user provides a detailed brief. Options to offer: (1) recall vault, (2) NotebookLM, (3) both, (4) skip. When a literature folder, a vault or a NotebookLM notebook is at hand, also offer (5) a wording profile, even if the triggers above do not apply: an outlet profile of the conference and a topic profile for the terms (d3-scientific-wording, Step 1b). Print: [HH:MM:SS] ENRICHMENT: [skipped|recall|notebooklm|both] [+ wording profile].
Resolve save location and create folder (see Project setup section above).
Initialize progress.md with phase-by-phase timestamps.
Wording brief. Invoke d3-scientific-wording (Skill tool: d3-scientific-wording:d3-scientific-wording; when it is not listed, read <wording-dir>/SKILL.md). Genre poster; the outlet is the conference, the reader its attendees, the field comes from the paper. Ask together with the outline approval of Pass 0, not in a separate round: "Wording: standard scientific (default) or Simplified Technical English (ASD-STE100: at most 25 words per sentence, one meaning per word; helps a mixed-language audience)?" Record the brief under ## Wording in progress.md. Print: [HH:MM:SS] WORDING: poster, [venue], [standard|STE].

Phase 2: Multi-pass build

Three passes. Never skip a pass.

Pass 0: Content outline (MANDATORY)

If a paper was scanned, use the extracted content as primary input.
If a knowledge brief exists, layer it in.
Draft a bullet-point outline (Title, Authors, Motivation, Problem, Approach, Results, Literature in the footer). For each: what goes in it.
Run the glance test — title + box headers + headline number must telegraph the contribution at 2m.
Present to user and get approval.
Print: [HH:MM:SS] OUTLINE: Approved by user.

Do NOT write any LaTeX until the outline is approved.

Pass 1: Skeleton

Create poster.tex using the Document Skeleton above.
Fill in all \headerbox{Title}{...} entries matching the approved outline.
Add % TODO: [description] placeholders inside each box.
Compile — must produce a 1-page A0 PDF.
Print: [HH:MM:SS] SKELETON: 1 page A0, [N] boxes, compiles clean.

Pass 2: Fill + compile

Fill each box content one at a time, in order: Motivation → Problem → Approach → Empirical Results.
Compile after each box is filled — verify the box height hasn't pushed neighbours into clipping.
Remove all % TODO: comments.
Final compile + visual validation (see Compilation and Validation section below).
Print: [HH:MM:SS] FILLED: All boxes complete, 1-page A0.

Content development

Define the core message

Audience: conference attendees walking by (30-60s for overview, 2-5 min if they stop)
Purpose: should a passing reader recognise the contribution and consider asking about it?
One-sentence summary: if the reader remembers one thing, what is it?

Plan the layout

Use the inverted pyramid + glance test: a viewer 2m away should grasp the topic from the title + box headers alone.

Content outline (MANDATORY before LaTeX)

Draft a bullet outline and get user approval. Example:

Title: [Paper title, shortened if needed]
Authors: [1-2 names + affiliation + email]
Motivation:
  - 2-3 sentences: problem framing + headline contribution
  - 1 \highlight{} on the key claim
Problem:
  - Formal setup (1-2 equations max)
  - keypoint: one-sentence takeaway
Approach:
  - 2-3 method paragraphs (bold lead-ins: Prescriptive analytics, Global training, Rolling horizon)
  - 1 figure (method overview, no caption)
Empirical Results:
  - 2-3 infoboxes (headline metrics)
  - 1 comparison table (booktabs, NeurIPS style)
  - 1-2 sentence conclusion with \highlight{} on the win margin
Literature (footer References slot):
  - 3 references

Budget content space (concise by default)

Posters are not papers and not one-pagers. A passer-by reads the title and headline number, stops if interested, and grasps the contribution from 3-5 sentences plus a figure. Anything more is wasted.

Default budgets for the 2-column layout:

Box	Default capacity
Motivation (col 0, top)	60-100 words, 1 `\highlight{}`, NO figure
Problem (col 0, bottom)	80-150 words + 1-2 equations + 1 `\keypoint{}`
Approach (col 1, top)	100-180 words (2-3 bold-lead paragraphs) + 1 figure
Empirical Results (col 1, bottom)	30-60 words + 2-3 infoboxes + 1 comparison table
Footer References slot	3 short references (authors, year, venue); 5 at most

If you exceed these by more than ~20%, you are overpacking. Cut sentences before adding line-spacing tweaks.

Wording. Every sentence on the poster follows d3-scientific-wording (<wording-dir>/references/rules.md; with STE also <wording-dir>/references/ste.md): direct claims instead of "not X but Y", related work reported neutrally, claims within the evidence, one term per concept. The rules below are the poster's form on top of them.

Concise writing rules (apply during Pass 2):

No throat-clearing. Cut "It is important to note that", "Our results suggest that", "We propose a novel approach that..."
Verb-first bullets. "Achieves 30% cost reduction" not "We achieve a 30% cost reduction".
Numbers before prose. Put 30% in an infobox and skip the prose version.
One sentence per idea. No compound sentences with subordinate clauses.
No background the venue audience already knows. A NeurIPS poster doesn't need to explain what regret is.
No "future work" section. Posters describe finished work.

Compilation and validation

Step 1: Compile locally (XeLaTeX preferred)

From poster/drafts/:

xelatex -interaction=nonstopmode poster.tex

With \cite{} + references.bib:

xelatex -interaction=nonstopmode poster.tex
bibtex  poster
xelatex -interaction=nonstopmode poster.tex
xelatex -interaction=nonstopmode poster.tex

Engine notes: XeLaTeX is default (loads Inclusive Sans UW via fontspec). LuaLaTeX works the same. pdfLaTeX falls back to a Helvetica clone (helvet) — use only if both Xe and Lua fail.

Step 1b: Overleaf fallback (only if local fails AND submodule path is available)

When local fails with a missing-package / missing-font / environment error you can't resolve, fall back to Overleaf. Only applies when the save location is inside a paper submodule.

Lab convention: paper submodules are on GitHub, Overleaf is connected via GitHub Sync. Flow: push to GitHub → Overleaf pulls → compile in browser.

cd <submodule-path>
git remote -v   # expect github.com/<org>/<paper-repo>.git, NOT git.overleaf.com
git add poster/
git commit -m "Add D3 poster"
git push

Then tell the user:

Open the connected Overleaf project
Menu → GitHub → Pull GitHub changes (or wait for auto-sync)
Menu → Settings → Main document: poster/drafts/poster.tex
Menu → Settings → Compiler: XeLaTeX
Click Recompile

Once it compiles, download the PDF and place at poster/drafts/poster.pdf so the rest of validation still applies. Don't forget to update the outer repo's submodule pointer:

cd <outer-repo-root>
git add <submodule-path>
git commit -m "Bump paper submodule with poster"

When local compile succeeds, do not mention the Overleaf path — it adds noise.

Step 2: Visual validation (MANDATORY)

A0 PDFs are huge. Render with uv run --with pymupdf to keep dependencies isolated:

mkdir -p review/
uv run --with pymupdf python -c "
import fitz
doc = fitz.open('poster.pdf')
for i, page in enumerate(doc):
    pix = page.get_pixmap(dpi=150)
    pix.save(f'review/page-{i+1:03d}.png')
    print(f'Page {i+1}: {pix.width}x{pix.height}')
doc.close()
"

Note on uv: this skill uses uv run --with <pkg> for all Python — never install packages globally.

Inspect each PNG with the Read tool. Check for: content clipping; header (navy box, white university logo, title and authors fully inside, blue seal flush with the box's right and bottom edges); section bars (#ECECEC with navy text, cornflower with white text only where emphasis is set); footer (navy rule, D3 logo first in the logo row, references inside the right half); font (Inclusive Sans UW, not a fallback); colors (D3 palette only, d3-orange only for highlights and alerts); TikZ and figures fit inside parent boxes.

Step 3: Check compilation warnings

grep -c "Overfull\|Warning" poster.log

A handful of underfull-hbox warnings is expected. Investigate Overfull \hbox exceeding 50pt (content too wide), Font ... not found (use XeLaTeX), File ... not found (missing logo / figure / baposter.cls), and d3-poster Warning: The title block was set at ... or The references block ... (shorten that text). Also run grep "Missing character" poster.log | grep -v nullfont: any hit is a character Inclusive Sans UW lacks (see references/error-handling.md #7). The ; lines in nullfont come from baposter.cls and are harmless.

Step 4: Lint the wording

Lint the wording of the source (add --ste when the user chose Simplified Technical English); fix every error and decide every warning:

uv run <wording-dir>/scripts/check_wording.py poster.tex --genre poster

Overflow & clipping

baposter does NOT push overflow to a new page like the article class — it silently clips. This is the #1 cause of broken posters. There's no LaTeX error; you only see it in the rendered PNG.

Prevention

Budget box content BEFORE writing LaTeX (see Content development → Budget content space).
Compile after each box is filled — catch clipping early.
Shorten text, don't shrink fonts — baposter sets its own scale per a0paper; manually scaling text breaks consistency.
One equation per box (max two) — equations consume a lot of vertical space at poster scale.
Lists: max 5-6 items per box.
Tables: max 6 rows at default scale.
One \keypoint{} per box max.

Detection checklist

Output is exactly 1 page (extra pages = layout collapse)
No box body content clipped at its bottom edge
No infobox/card/figure overflows its parent box
All \keypoint{} callouts fully render
Title and author lines fully inside the navy header box (no "set at ... of its normal size" warning)
No "Overfull \hbox" warnings > 50pt in the .log

If clipping detected

Identify the overfull box (visually scan the rendered PNG).
Cut the lowest-priority content in that box.
If still clipping: shorten lists (max 4 items), reduce table rows, compress prose.
If a \keypoint{} causes clipping: move it earlier or shorten its text.
Never shrink fonts. Cut content.
Recompile and re-render.

Review

Review depth. Three levels; the user picks one in words, otherwise use standard and say so in one line ("Review: standard. Say 'light' to skip the reviewer agents.").

Depth	What runs	Token cost
`light`	The automated checks (compile log, page count, clipping checklist) and your own look at the rendered pages. No reviewer agents.	almost none
`standard` (default)	The two reviewer agents below, in parallel, one round. Apply their fixes, recompile, rerun the automated checks. No second review.	about 100k tokens, largely independent of document length
`full`	The two reviewers in a loop until both scores are at least 9, at most 4 rounds; then hand the remaining findings to the user.	two to four times `standard`

"quick", "no review", "light", "short on tokens", "cheap" mean light. "thorough", "full review", "submission-ready" mean full. In light, replace the reviewers by one pass of your own over both rubrics and list what you checked.

For standard and full, load references/dual-agent-review.md: it has the two reviewer prompts (Content, Visual Design), the orchestrator summary template, and the score-to-README protocol. Dispatch both agents in parallel (single message, two Agent tool calls with subagent_type=Explore); both inspect the rendered PNG. Surface their scores and feedback to the user in chat: what was checked, what passed, what failed, what was changed. standard stops after one round of fixes; full iterates until both Overall scores are at least 9, or 4 rounds elapse.

Editing & versioning

When editing an EXISTING poster, use git, not filename prefixes. The working file is always poster.tex and produces poster.pdf — stem-match invariant.

Detect editing mode. Print [HH:MM:SS] EDITING MODE: Found existing poster -- [filename].
Snapshot via git BEFORE editing:
```
cd <repo-with-the-poster>
git add poster/drafts/poster.tex && git commit -m "Pre-edit snapshot"
```
If git isn't available (rare): cp poster.tex .poster.tex.pre-$(date +%s). Git is strongly preferred.
Read the existing file — understand box structure, content, style.
Apply edits in place to poster.tex. Single working file; do not create v2_poster.tex.

Document changes in revision_notes.md:

## Revision YYYY-MM-DD HH:MM:SS
- Replaced placeholder figure with method-overview.pdf
- Updated Empirical Results table with final numbers
- Resulting content score: 9/10, visual score: 9/10

Compile and validate — see Compilation and validation above.

Wording corrections from the user are logged verbatim under ## Wording in progress.md, and each stated fix is applied everywhere it recurs on the poster.

Promotion to final/: before copying, ask the watermark question of d3-scientific-wording (Step 5) once; clean poster.tex only on a yes, then recompile and rerun the checks. When the user approves, copy poster/drafts/{poster.tex, poster.pdf, baposter.cls, assets/} into poster/final/ with the SAME filenames. Write final/README.md capturing the date, short title, venue, and final review scores.

Macros & styling — at a glance

The most-used macros for a typical poster. See references/macros-and-styling.md for the full inventory (more cards, font-size table, full color palette, table conventions, title/subtitle syntax).

Macro	Use
`\dthreelogos`	D3 logo across the first two cells of the footer logo row (5th argument of `\begin{poster}`); the university logo is always in the header
`\dthreelogocell{file}`	One more logo cell in the footer row (partner or funder logo), after `\dthreelogos`
`\dthreereferences{...}`	Footer References slot (right half): 3-5 short references
`\highlight{text}`	Inline d3-orange bold; 2-3 per box max
`\keypoint{text}`	d3-navy tinted callout box with left accent; 1 per box max
`\headerbox{Title}{keys}{content}`	Primary content box (baposter built-in)
`\hhrule`	Horizontal d3-navy rule (separator inside a box)

TikZ styles available inside tikzpicture: card, card gray, card sand, card highlight, card alert, takeaway, infobox, tag. Box key: emphasis (cornflower section bar with white text).

D3 colors (use ONLY these): d3-navy (#02246C, university navy), d3-orange (#F29100, D3 accent), d3-gray, d3-lightblue, d3-darkblue, plus d3-blue, d3-sky, d3-sand, d3-teal, d3-yellow, d3-berry, d3-hairline, d3-headergray (full table in references/macros-and-styling.md). Never red/blue/green or arbitrary hex codes.

Error handling

Load references/error-handling.md when a compile error or rendering issue appears. Known issues:

Class file not found — baposter.cls must sit next to poster.tex, not under assets/.
Font not found (Inclusive Sans UW) — use XeLaTeX, verify assets/fonts/ exists.
\d already defined (siunitx conflict) — d3-poster.sty does \let\d\relax before loading siunitx; outdated .sty means recopy.
Box headers with a gradient instead of the solid fill: baposter's default headershade=shadeLR is a gradient. Keep headershade=plain (already in Document Skeleton).
Calibri font error — using unpatched upstream baposter.cls. Use the vendored copy at the skill root.
Box content clipped (no LaTeX error) — baposter silently clips overflow. See Overflow & clipping above.
Greek letters or symbols vanish — Inclusive Sans UW has no Greek and lacks ≤ ≥ ≈ ≠ ∞ √ ∈ ∑ ∫ ∂ ⇒ ′ ″ ✓; d3-poster.sty maps these to math mode. Other missing glyphs: write them in math mode.
Title or references set smaller — the block did not fit its slot; shorten it.

Quality checklist

Before marking a poster complete:

Files & structure

baposter.cls at poster/drafts/ root, next to poster.tex
assets/d3-poster.sty, assets/fonts/, assets/logos/ present
No leftover example-poster.tex / example-poster.pdf in the project
PDF compiles without errors using XeLaTeX
PDF is exactly 1 A0 page
poster.pdf stem matches poster.tex

Versioning (if editing)

Pre-edit snapshot committed via git
Single poster.tex edited in place — no v2_/v3_ filenames
revision_notes.md updated with a dated entry

Content

Visual validation

PDF rendered to PNG and inspected
No box content clipped at any edge
Header: navy box, white university logo, title fully visible, blue seal flush with the box's right and bottom edges
Footer: navy rule, D3 logo first in the logo row, references inside the right half
Inclusive Sans UW renders correctly (not a fallback); no missing Greek letters or symbols
Colors: section bars #ECECEC with navy text (cornflower only for emphasis), d3-orange for highlights only
All figures and tables fit within parent boxes
No "Overfull \hbox" warnings > 50pt

Review

Review done at the chosen depth: light checks and own pass listed; standard one reviewer round applied; full Content and Visual scores ≥ 9/10
final/README.md written with date, title, venue, scores

Decision points

Ask the user when

Input tier (A/B/C/D) — first question, before everything else
Paper source — if multiple paper candidates exist
Column count — default 2; confirm before assuming 3
Title & subtitle — title confirmation; subtitle off by default, ask only if mentioned
Author block — 1-2 names + affiliation + email on one line
Headline metric — the single number a viewer should remember (goes in an infobox)
Figure preference — which 1-2 figures from the paper to feature
Architecture figure — if detected: use-as-is / simplify / replace / skip (see references/extraction.md §8)
Captions on figures — off by default; ask only if mentioned (e.g., venue requirement)
Bibliography filtering — which 3-5 references to feature in the footer References slot
Knowledge enrichment — vault recall / NotebookLM / both / skip (only when the request is high-level)
Review depth — do not ask; use standard unless the user's words pick light or full, and say which one runs
Wording mode — standard scientific wording (default) or Simplified Technical English; asked with the outline approval
Invisible marks — the watermark question of d3-scientific-wording, once, before promotion to final/

Decide independently

Which \headerbox keys to use for positioning (within the chosen column count)
Which TikZ macro fits each piece of content (infobox vs. card vs. card alert)
Spacing adjustments within a box (\vspace)
Default 2-column layout (Motivation+Problem | Approach+Results) unless user requests otherwise

Ask in your favorite AI

Documentation