ResearchStudio-Reel turns one paper PDF into an editable poster, talk video, and bilingual blog by sharing a single upstream extraction across five Claude Code skills and gating each generator on hard categorical render checks instead of soft VLM-as-judge scores.
You’ve shipped a product that takes a source document and generates three downstream artifacts (say, a spec, a demo video, and a marketing page). Today each artifact is its own pipeline that re-parses the source, ships a flat PDF or MP4 the author cannot reopen, and passes QA when a learned preference score crosses some threshold, even if half the page reads empty. This paper argues that pattern is architecturally wrong. Prior paper-to-artifact systems like Paper2Poster each re-extract figures, captions, and metadata inside their own renderer, so cross-artifact consistency (the figure on the poster matching the one in the blog) falls on the user.
The pipeline is five skills sharing one data bundle. Paper2Assets reads the PDF once and emits body text, cleaned figure crops, a per-figure manifest, metadata, logos, QR codes, and a nine-section structured summary. Three generators (Paper2Poster, Paper2Video, Paper2Blog) consume that bundle verbatim; Paper2Reel binds their outputs into an interactive HTML viewer.
The load-bearing mechanism is what the authors call a measured-fill loop, most explicit in the poster skill. Instead of scoring layout on a 1-10 aesthetic scale and stopping at a plateau, each section’s painted-content height divided by its card height is quantized into five categorical verdicts: EMPTY, SPARSE, FULL, SPILLAGE, OVERFLOW. Only FULL (0.90 to 1.00) passes. Each verdict maps to a specific edit (append reserved paragraph, polish up, polish down, drop a section), sized by the signed pixel delta the browser measurement reports.
while not all_sections_full(poster):
render_headless(poster)
for section in poster.sections:
ratio = section.content_h / section.card_h
verdict = quantize(ratio) # EMPTY|SPARSE|FULL|SPILLAGE|OVERFLOW
if verdict != FULL and not already_tried(section, verdict):
apply_move(section, verdict, pixel_delta)
if round_counter.tripped(): ship_best_measured(); break
render_to_pdf_png_pptx(poster)
Editability is preserved by reconstructing PowerPoint shapes from the live DOM (walking node geometry via getBoundingClientRect, mapping CSS to native OOXML text frames, bullets, equations, rounded rectangles) rather than rasterizing the finished page. The video skill delegates deck authoring to ppt-master and adds a timeline sidecar so section IDs, audio windows, subtitles, and slide frames stay addressable after export. The blog skill writes two Word files from one shared evidence map, gated by a layout check that catches near-blank pages and orphan lines.
The prevailing approach in paper-to-artifact automation is to grade output with a continuous VLM-as-judge score and accept whatever plateaus above threshold. This paper shows the opposite. Quality gates should be categorical pass/fail render checks tied to measurable geometry, not learned preference scores, because a discrete-edit loop oscillates in the last few percent and a soft score cannot tell you which specific section is broken. The cleanest evidence is the ablation where the same base model, prompted once versus wrapped in the measured-fill loop, jumps roughly a full point on Layout under identical judges.
The load-bearing finding is the model-held-fixed ablation. Prompting claude-opus-4.8 once for a full poster yields an aesthetic mean of 2.76 and Layout of 2.83; wrapping the same model in the skill composition and measured-fill loop lifts these to 3.52 and 3.97. The same substitution under gpt-5.5 moves aesthetics from 3.03 to 3.36. The gain is attributable to the loop and the shared extractor, not to a stronger generator.
•
On the 100-paper Paper2Poster benchmark, the system leads every aesthetic and information sub-criterion against prior automated poster systems and single-shot frontier LLMs, under two held-out judges.
•
It surpasses the authors’ own posters on aesthetics (3.52 vs 2.94) and information (3.90 vs 3.63), winning the overall score on 84 to 93% of papers.
•
PaperQuiz accuracy inverts: text-dense systems like P2P win comprehension while losing visually, exposing a genuine tension between raw text coverage and legible layout.
•
Full pipeline cost: about 89 minutes and 2.6M input / 276K output tokens per paper, with most traffic being cached-context re-reads billed at roughly 10% of fresh input.
•
Capability audit: it is the only system that ships all three editable artifacts (PPTX, MP4, DOCX) from one run.
Reach for this pattern when you have a pipeline generating multiple downstream artifacts from one source and you’re currently grading each with a preference score. Split the source extraction into a single shared owner that emits stable IDs (section, figure, claim) that every downstream reads verbatim. Replace the preference score with a categorical verdict tied to something you can measure deterministically (rendered geometry, file structure, playable audio, non-empty subtitle sidecar). The Measured-fill loop control flow (quantize, select move by verdict, size by measured delta, refuse to re-apply a move that already overshot) is the reusable primitive.
Code is released under MIT at aka.ms/ResearchStudio, with each skill’s SKILL.md as the workflow source of truth. System dependencies are poppler-utils, libreoffice, ffmpeg, headless Chromium, and Edge TTS for narration. The DOM-to-PPTX bridge is standalone-useful for any tool that wants an editable PowerPoint from an HTML document. The Paper2Poster benchmark and PaperQuiz rubric are reused verbatim for scoring.
When a discrete-edit loop keeps oscillating, quantize the target into named bands and let the verdict pick the move. Learned preference scores tell you the output isn’t good enough; they don’t tell you which section to edit or by how many pixels. A five-band categorical verdict tied to measured geometry does both, and it gives you a fixed-point termination condition instead of a plateau.
•
Quantitative evaluation only covers the poster; the video and blog are compared on capability checkboxes because no graded benchmark exists for either. The claim that categorical gates beat soft scores is proven for one of three artifacts.
•
The pipeline reuses only figures already in the paper, so it cannot draw the bespoke method diagram a human designer adds. Closing that gap reintroduces the hallucination risk the categorical gates were built to suppress.
•
Calibrated on ML, CV, and NLP venues where poster and figure conventions are uniform. Transfer to biomedicine or physics is untested, and each skill’s move catalogue may need extending per venue family.