Paper-to-Presentation Pipeline
Turns a batch of academic papers into a rigorous, fully traceable course slide deck you can actually present.
The skill only reads/writes files within a local project directory (PDF inputs, intermediates, .pptx output); no network calls, no credential access, no destructive defaults; data flow is transparent and provenance is traceable. Deducted for: no rollback mechanism, no isolation/confirmation constraints over dependent skills (pdf/pptx/ppt-creator), and unverified publisher identity.
Instructions are self-consistent with clear staging, abnormal-input handling (file-lock PermissionError fallback, figure-mismatch fallback, text-overflow prevention) and explicit failure markers (FIGURE-MISMATCH). Deducted for: no executed validation, no tests, and rules (DPI thresholds, collision checks) that depend on actual agent execution quality rather than mechanisms.
Triggers are explicit in both Chinese and English, audience is defined, 'Not For' states non-fit boundaries, and environment assumptions (Windows fonts) are declared. Deducted for: unverified assumption that required sub-skills ship with Claude Code, no version/environment compatibility notes, and partially semantic trigger boundaries.
SKILL.md and README are well-layered with directory conventions, thorough design rationale, and a clear MIT license. Deducted for: no versioning/changelog, no FAQ, no known-limitations section, and unclear ownership/maintenance/update path with an unverified publisher.
The claimed output (a provenance-traceable deck) is plausibly designed, and the verification gates show genuine quality consideration. Deducted for: static review cannot confirm outputs are directly usable, no sample outputs or third-party usage evidence, and comparative benefit over manual workflows is inferred only.
The repo contains only instructional files — no tests, no CI, no execution records or independent corroboration. All quality claims (readability thresholds, collision rules) are author assertions with no fact/inference separation.
- This is a static review only; no pipeline stage was executed and actual output quality is unverified.
- Availability of the pdf, ppt-creator and pptx sub-skills is assumed but unconfirmed; the pipeline fails without them.
- Verification gates and layout rules rely on agent compliance, not hard mechanisms, so reproduction is not guaranteed.
- Publisher identity is unverified and there is no versioning or changelog, so long-term maintenance is unclear.
What does this skill do, and when should you use it?
This is a Claude Code skill that transforms N paper PDFs into a 15–18 slide, ~20-minute course presentation through four sequential stages. Each stage produces verifiable intermediate artifacts, so failures only require re-running one stage. It enforces full claim provenance, locks the Ocean Gradient palette, requires every headline to be a complete assertion, and puts all citations in speaker notes. Built-in figure verification, a keyword-overlap guardrail, layout collision detection, and a readability check with automatic slide splitting catch the classic failure modes of AI-generated decks.
Reads PDFs from papers/ and runs: agent1 extracts all figures via the pdf skill (≥150 dpi, no content alteration) into an indexed inventory with section context and literal visual descriptions; agent2 extracts conclusive claims, quantitative results, and key metrics tagged to page/paragraph; agent3 applies the pyramid principle to produce a 10-field per-slide spec and runs a mandatory visual verification gate ([VERIFY] lines) on every referenced figure; agent4 generates final_presentation.pptx and speaker_notes.md page-by-page with pptxgenjs, applying a keyword-overlap guardrail, bounding-box collision detection, text-overflow estimation, aspect-ratio-aware sizing, and an effective-DPI/edge-density readability check that auto-splits slides with unreadable figures into a text page plus a full-figure page.
- An undergraduate or graduate student preparing a 15–20 minute course talk from papers related to their course
- A grad student giving a lab-group paper report that needs traceable citations
- Converting a conference poster into an oral presentation
- Visualizing a review article as a chapter-structured explanatory deck
- Presenters frustrated by garish AI palettes, vacuous headers, and inconsistent citations who want hard constraints
What are this skill's strengths and limitations?
- 100% of claims trace to source papers, cited to page/paragraph in speaker notes
- Four stages with independent verification gates—re-run one stage, not the whole pipeline
- Locked palette and assertion headlines eliminate AI slop colors and empty section titles
- Three-layer figure defense: visual verification gate, keyword-overlap guardrail, readability check with auto-split
- Explicitly documents non-use cases (pitch decks, non-academic content, single-paper lightning talks)
- Strict four-stage sequencing is overkill for a quick single-paper task
- Assumes an audience of STEM undergraduates past the intro course—wrong fit for lay audiences
- Auto-splitting unreadable figures adds a slide per flagged figure
- No test suite or sample outputs in the repo; quality depends on upstream Claude Code built-in skills
- Fonts guaranteed only for Windows; behavior on other platforms is unverified
How do you install this skill?
Run git clone https://github.com/elinglijiaoqiao/presentation-ppt-maker.git, then ln -s "$(pwd)/presentation-ppt-maker" ~/.claude/skills/paper-to-presentation; or simply copy SKILL.md into ~/.claude/skills/paper-to-presentation/. The pdf, ppt-creator, and pptx skills ship with Claude Code—no extra installation; optionally pair with a paper-search CLI for upstream retrieval (the source does not document how to install it).
How do you use this skill?
1) Create a project directory and place paper PDFs in papers/; 2) execute agent1 → agent2 → agent3 → agent4 in strict order, verifying each output before proceeding; 3) trigger phrases include "presentation from papers", "论文 PPT", "做 pre", or "课程汇报"; 4) collect final_presentation.pptx and speaker_notes.md.