Paper2ppt

Paper2PPT

Non-negotiable output

• Create an editable .pptx when authoring exists; never stop at an outline.
• Default to Simplified Chinese, 12-16 slides, and 15-20 minutes; adapt as requested.
• Add speaker notes and source labels to every result slide.
• Do not fabricate claims, values, mechanisms, methods, figures, or citations.

Preflight

Resolve <SKILL_DIR> to this loaded SKILL.md's directory: Claude Code substitutes ${CLAUDE_SKILL_DIR}; Codex uses the absolute path in its available-skills entry. Never locate helpers from the current working directory. Before analysis, run python "<SKILL_DIR>/scripts/preflight.py" INPUT --output PATH --json; add --create-output deliberately. Use --native-authoring or --native-pdf only after confirming equivalent authoring or extraction; neither bypasses source/output failures. Verify audience, duration, language, template, and source sufficiency.

Source sufficiency

Classify source sufficiency as full-paper, partial-paper, or notes-only; record it in qa_report.md.

Never imply unsupported full-paper knowledge. Reconstruct design, values, mechanisms, or diagrams only when every component and relationship is source-supported; otherwise omit. Label reconstructions and distinguish them from an original paper figure.

Core workflow

1. Read the evidence-first workflow when generating from source, extracting figures, or doing substantive build or revision.
2. Extract only source-supported questions, gaps, design, claims, evidence, validation, limitations, and implications.
3. Build and use a claim-to-source map before drafting slides.
4. Classify the paper. Read the paper-type routes when the user has not fixed the narrative or slide plan and the task is not a narrow repair; otherwise follow the supplied structure or repair scope.
5. Plan one claim per slide; build with notes and attribution, then run the quality gates.

For a narrow existing-deck audit or repair, load only guidance relevant to the requested change.

Evidence and attribution

• Treat figures as evidence, not decoration.
• Preserve figure labels, scales, colorbars, statistics, and qualifiers.
• Never use AI-generated imagery as data evidence.
• Give each result slide a conclusion-style title, one claim, a source label, and evidence/limit notes.
• Trace assets by source location, figure or table, panel, crop status, and destination slide.

Read adaptive slide design only when designing or revising layouts or correcting visual defects.

Tool routing

Use available presentation tools; otherwise use PyMuPDF/Pillow and python-pptx. Install only when permitted.

After preflight, read tool routing when selecting or falling back between extraction, authoring, notes, or rendering tools.

Quality gates

Run all available gates before delivery:

1. Content: verify sources, numbers, labels, marked inferences, and source sufficiency.
2. Structural: reopen the PPTX; verify editability, media, notes, sources, and files.
3. Visual: render every slide when possible; inspect clipping, overlap, readability, contrast, and figures. Otherwise record unavailable visual QA in qa_report.md.

Use exact states: pass only when all three gates pass; structurally-verified-with-disclosure only when content and structural QA pass and visual QA is unavailable solely because no renderer exists; blocked for unresolved content/structure, failed editable authoring/reopening, or unresolved visual failure when a renderer is available.

Read quality gates before final verification or when fixing QA failures.

Output contract

Create only:

• presentation_zh.pptx by default, or presentation_<language-code>.pptx when the user explicitly requests another language;
• qa_report.md;
• asset_manifest.md and assets/ only when assets are extracted.

Do not create PDF, XLSX, outline, script, or source-map deliverables by default.

Failure handling

• On access failure, request a local PDF; never bypass access controls.
• Use OCR only after ordinary text extraction fails.
• Narrow claims and mark uncertainty for partial material.
• If an optional template is invalid, use a neutral academic layout; stop only when the template is mandatory.
• Without editable authoring, fail before extensive analysis and name the missing capability.

Completion checklist

• Confirm the PPTX opens and matches language, duration, and audience.
• Record sufficiency and claim-to-source outcomes.
• Verify each result slide's claim, source, figure integrity, and notes.
• Complete all available QA; disclose unresolved limits in qa_report.md.
• Deliver only the minimal outputs.