Swapping PPTX Templates

Swapping PPTX Templates

Overview

Moving a deck's content into a different template's look. PowerPoint has no "reflow into another template" operation, so this is a slide-by-slide rebuild in raw OOXML. This skill is the judgment layer on top of the mechanical workflow.

REQUIRED BACKGROUND: Use the official pptx skill (Anthropic document-skills plugin) for the unpack → add_slide → edit → clean → pack mechanics and its scripts. This skill covers the four things that skill does NOT warn about, each of which caused a real defect.

The four non-obvious rules

1. Redesign, don't relocate

A "template" usually ships 3-5 generic layouts (title / banded-header / framed-card / blank), not equivalents of the source's custom diagrams. Copy-pasting shapes drags the source's colors and geometry into the new deck. Instead: extract each slide's content (text, data, image refs) and rebuild the visual from scratch in the template's palette, using its layouts only as background/chrome. Match content type to the closest layout; layer plain text boxes + shapes on top for anything beyond title+body.

2. Never hand-type CJK (or any non-ASCII) as numeric XML entities

Deriving 住... decimal codepoints from memory for Chinese/Japanese/Korean text is the single highest-risk step. Each wrong codepoint decodes to a valid but wrong character (圈→圭, 驗→鰓, 輪→載), so pack/validate never errors and the corruption is silent.

• Get ground truth: python -m markitdown source.pptx > source_text.md
• Insert CJK by script substitution from that file (str.replace), never by typing entities or characters into an edit by hand.
• Before packing, run the audit (path in Claude Code shown; elsewhere use the repo path):

AUDIT="$CLAUDE_PLUGIN_ROOT/skills/swapping-pptx-templates/scripts/audit_cjk.py"
python "$AUDIT" source_text.md <build_dir>   # exit 1 = suspicious chars, stop and fix

It flags any CJK char in the build absent from the source (near-zero false positives).

3. Template backgrounds are hazard zones

A template's p:bg is often a rasterized blipFill, not theme colors — so it can contain baked-in text (a tagline) and a logo that your content must avoid. Detect their bounds programmatically, don't eyeball a thumbnail:

from PIL import Image
im = Image.open("ppt/media/image1.PNG").convert("RGB")  # the bg blip
# scan for tagline/logo pixels vs the known flat bg color; get px bbox, then:
# EMU_per_px = slide_width_emu / image_width_px   # e.g. 12192000/1280 = 9525

Convert the bbox to EMU and treat it as a keep-clear rectangle for every slide on that bg.

4. Visual QA is an adversarial pass by fresh eyes

After packing: soffice --headless --convert-to pdf out.pptx && pdftoppm -jpeg -r 150 out.pdf slide. Then dispatch a subagent (no authoring context) told to find problems — overlaps with the bg watermark/logo, text overflow, low-contrast label/fill pairings, lopsided whitespace. You will not catch these yourself; you've been staring at the coordinates. Fix and re-render until a pass is clean.

Red flags — STOP

• About to type a Chinese/Japanese/Korean character or &#NNNN; entity into an edit → don't; substitute from the markitdown extraction instead.
• Packing without having run audit_cjk.py → run it first.
• Placing content in the top-right or a bottom band without checking the bg image's logo/tagline bbox.
• Declaring done after only self-review of the rendered slides → dispatch the fresh-eyes QA subagent.

Common mistakes