Thesis Defense PPTX Agent Skill

Thesis Defense PPTX

Use this skill for local, editable .pptx thesis defense decks that must follow a supplied university/lab/company PowerPoint template.

Operating Principles

• Platform: Windows + Microsoft PowerPoint for COM-driven cloning, export, and overflow inspection. On macOS/Linux, use the python-pptx-only parts and skip COM-based quality gates.
• Treat the visual template as the source of truth. Preserve its cover, colors, fonts, navigation, card styles, logos, and slide proportions unless the user explicitly asks to redesign.
• Read the thesis source first: PDF, LaTeX, figures, old PPT, experiment scripts, tables, and captions. Do not generate from topic/title alone when source files exist.
• Build a source-asset inventory before placing images: extract \includegraphics entries and captions from the final TeX/PDF, then use only thesis-cited figures or images the user explicitly approves. Do not carry over pictures from a template, old progress deck, or draft figure folder merely because they look relevant.
• Prefer copying native template slides and replacing content over rebuilding from blank slides.
• Keep slides concise and presentation-oriented. Convert thesis prose into defense talking points.
• Preserve the template's font sizes and hierarchy by default. For projection readability, first shorten text, split dense slides, or use the template's larger existing layouts; only change font sizes locally when the template itself provides no readable fit.
• Always generate an editable .pptx, not only an outline, image deck, PDF, or HTML.
• Always run a visual quality gate before final response: export slide images, inspect a contact sheet plus key full-size slide screenshots, check text overflow, check stale template words/images, and verify the PPTX opens.

Recommended Workflow

• Identify inputs:
• thesis project path or full thesis document
• visual template PPTX
• optional old defense/report PPTX
• output directory and desired filename

• Extract thesis context:
• Use scripts/extract_thesis_context.py to summarize PDF/TeX text and list candidate figures.
• Read source files directly when details matter: abstract, introduction, method, experiments, results, conclusion.
• Create an explicit list of final-thesis figures from \includegraphics and captions. Treat this as the allow-list for PPT images unless the user requests new visuals.
• Collect defense-critical facts: problem, motivation, modeling, algorithms, experimental setup, baselines, metrics, key results, limitations, future work.

• Analyze the template:
• Export template slides to PNG if PowerPoint is available.
• Inspect cover, TOC, navigation, section, body/card, chart/image, summary, and closing slide styles.
• Note fonts, font sizes, red/brand color, card dimensions, line widths, and spacing.

• Build a native template skeleton:
• Use PowerPoint COM (scripts/clone_template_deck.ps1) when PowerPoint is installed.
• Copy or duplicate original template slides in the needed order.
• Use python-pptx only to replace content and add images/shapes after the skeleton is valid.

• Fill content:
• Use 12-16 slides for undergraduate defense unless user specifies otherwise.
• Typical order: cover, TOC, background, problem definition, work summary, system/method, modeling, implementation, experiment setup, training/results, comparison, robustness/multi-seed, innovation/limitations, summary/outlook, thanks.
• Use final-thesis figures first and verify each inserted picture against the source-asset inventory. If a picture is from a template or old deck and is not thesis-cited, replace it or remove it.
• Keep dense tables limited to headline metrics; move detailed prose into verbal notes or omit. Prefer one message per slide, large existing template placeholders, metric cards, and short bullets over shrinking text or adding small multi-column tables.
• Match the template's typography before changing sizes: preserve title/body/table/navigation sizing where possible, and use content reduction or slide splitting before manual font scaling.

• Quality gate:
• Export final PPTX to PNG with scripts/export_pptx_png.ps1.
• Create a contact sheet with scripts/make_contact_sheet.py.
• Run scripts/inspect_pptx_overflow.ps1.
• Run scripts/scan_pptx_text.py for slide count and stale template terms.
• Inspect the contact sheet manually, then open any risky slide full-size. COM overflow is not sufficient: also check visible clipping, text crossing card borders, small text caused by overfilling, stale template images, image aspect ratios, and text/card overlap.
• Check stale template text, placeholders, wrong names/dates, wrong navigation labels, red active-nav white text, and non-thesis images accidentally inherited from the template or old progress deck.
• Fix and repeat until the exported deck is clean.

• Final response:
• Give the full output path, total slide count, main changes, validation performed, and any unresolved manual confirmations.

Script Usage

Extract thesis context:

python scripts/extract_thesis_context.py --input "D:\path\to\thesis" --output "D:\path\to\thesis_context.md"

Clone template slides into a native skeleton:

powershell -NoProfile -ExecutionPolicy Bypass -File scripts/clone_template_deck.ps1 `
  -Template "D:\path\template.pptx" `
  -Output "D:\path\output.pptx" `
  -SlideSequence "1,2,3,4,4,6,7,8,9,10,11,12,13,14,15"

In this example slide 4 is deliberately reused as a sub-section page and slide 5 is skipped.

Export final deck:

powershell -NoProfile -ExecutionPolicy Bypass -File scripts/export_pptx_png.ps1 `
  -Pptx "D:\path\output.pptx" `
  -OutDir "D:\path\visual_check" `
  -Width 1600 -Height 900

Inspect possible overflow:

powershell -NoProfile -ExecutionPolicy Bypass -File scripts/inspect_pptx_overflow.ps1 `
  -Pptx "D:\path\output.pptx" `
  -Tolerance 40 `
  -OutputJson "D:\path\overflow.json"

Make a contact sheet:

python scripts/make_contact_sheet.py --input "D:\path\visual_check" --output "D:\path\contact_sheet.png"

Scan deck text and forbidden terms:

python scripts/scan_pptx_text.py `
  --pptx "D:\path\output.pptx" `
  --bad "中期检查,项目简介,系统框图,算法设计,成果展示,总结分析,TODO,占位"

For quality criteria, read references/pptx_quality_gate.md.

Iteration Tips

The fastest way to fail a content fill is guessing strings instead of reading them. Use this loop:

• Dump first. Run scripts/dump_pptx_content.py --pptx <skeleton> --output dump.md to list every shape's exact text, plus all table cells and picture names. Open dump.md and copy the strings you intend to replace.
• Prefer partial match for body text. Use replace_partial_text(slide, {"keyword": "new full text"}) instead of replace_exact_text whenever the body sentence might differ from the template by a stray space, half/full-width quote, or —— vs —. Pass min_len=20 to avoid short-key collisions with navigation tags.
• Keep replace_exact_text for short tags. Navigation labels and section titles ("项目简介", "工作总结") are stable; exact match is safer because it cannot accidentally rewrite a body paragraph that happens to contain the same word.
• Use the table/picture helpers. write_table(table, rows) preserves cell font/color/size, and replace_picture(slide, old_pic, new_path) swaps an image while keeping position and size — both avoid the boilerplate of removing and re-creating shapes.
• Re-run overflow + scan after every fill pass. Overflow caused by longer content is the single most common bug; fix by shortening copy first, then by manual \n line breaks, and only resize the textbox if the template visual really allows it.
• Encoding. All Python scripts shipped here force stdout to UTF-8, and the PowerShell scripts force the console output codepage to UTF-8, so cp936/gbk consoles do not corrupt JSON output containing CJK paths or Unicode symbols (−, Δ, ✓).

Script Usage

Dump every slide's shapes/text/tables/pictures (run this BEFORE constructing replacement dicts):

python scripts/dump_pptx_content.py --pptx "D:\path\skeleton.pptx" --output "D:\path\dump.md"

Add --slide 4,8,9 to limit to specific slides while iterating.

Implementation Notes

• Requires Python 3.10+ and the packages in requirements.txt.
• On Windows with Microsoft PowerPoint installed, PowerPoint COM is the most reliable way to duplicate template slides and inspect real rendering.
• python-pptx is good for editable text, shapes, images, tables, and simple charts, but it does not perfectly model every PowerPoint layout/rendering behavior.
• scripts/pptx_template_tools.py uses THESIS_PPTX_BRAND_RED when set (#830604, 830604, or 131,6,4) so new school templates do not require source edits for the primary red.
• When a PowerPoint add-in blocks COM automation, disable only the broken add-in autoload if necessary and report that action.
• Use backups before overwriting a deck.
• Never revert unrelated user files while iterating.