ppt-from-template-skill

内容来源：SKILL.md（标准 Skill 格式） · 原始地址 · 查看安装指南

原始内容

---
name: generating-pptx-from-template
description: "Generates fully populated PowerPoint presentations that inherit the exact colors, fonts, layouts, and branding from a user-supplied .pptx template. Use when the user uploads or references a .pptx template file and asks to create or generate a presentation on a given topic. Triggers on: use my template, create slides from my pptx, generate a deck using my template, fill in my template, make a presentation from the attached file, build slides using my brand template, or when the user shares a .pptx file and wants a polished output matching their brand. Also use when multiple .pptx files are uploaded and the user wants a presentation combining layouts from all of them."
license: MIT
---

# Generating PPTX from Template

Produces polished PowerPoint presentations using an uploaded `.pptx` file as the
design source of truth. The output inherits the template's exact visual identity
— no generic AI aesthetics, no hardcoded colors or fonts.

---

## Setup

Install required tools before any other step:

```bash
pip install "markitdown[pptx]" Pillow defusedxml --break-system-packages -q
which pdftoppm || apt-get install -y poppler-utils -q
```

All helper scripts live in two locations:

```
# Public pptx skill scripts (structural operations)
/mnt/skills/public/pptx/scripts/office/unpack.py
/mnt/skills/public/pptx/scripts/office/pack.py
/mnt/skills/public/pptx/scripts/office/soffice.py
/mnt/skills/public/pptx/scripts/add_slide.py
/mnt/skills/public/pptx/scripts/clean.py
/mnt/skills/public/pptx/scripts/thumbnail.py

# This skill's scripts (QA and analysis)
/mnt/skills/user/generating-pptx-from-template/scripts/audit_overflow.py
/mnt/skills/user/generating-pptx-from-template/scripts/check_placeholders.py
```

---

## Workflow Overview

```
Step 1  →  Ingest template (markitdown + thumbnails + unpack)
Step 2  →  Plan content (layout inventory + slide map)
Step 2.5 → Measure placeholder capacity BEFORE writing any content
Step 3  →  Structural edits (add/remove/reorder slides)
Step 3.5 → Populate content (XML editing with overflow prevention)
Step 4  →  QA (content check + visual render + overflow check)
Step 5  →  Deliver
```

---

## Step 1 — Ingest the Template

Locate and copy the uploaded template:

```bash
ls /mnt/user-data/uploads/*.pptx
cp /mnt/user-data/uploads/YOUR_TEMPLATE.pptx /home/claude/template.pptx
```

Extract all text and structure to understand existing content:

```bash
python -m markitdown /home/claude/template.pptx
```

Generate visual thumbnails — **this is mandatory, not optional**. You must see
the actual layout, colors, design motifs, and photo placement before writing
a single character of content:

```bash
python /mnt/skills/public/pptx/scripts/thumbnail.py \
  /home/claude/template.pptx /home/claude/thumbnails
```

Unpack to XML for editing:

```bash
python /mnt/skills/public/pptx/scripts/office/unpack.py \
  /home/claude/template.pptx /home/claude/unpacked/
```

After unpacking, run a quick structural survey:

```bash
# How many slides?
ls /home/claude/unpacked/ppt/slides/slide*.xml | grep -v _rels

# What layouts does each slide use?
for f in /home/claude/unpacked/ppt/slides/slide*.xml; do
  echo "$f:"; grep -o 'slideLayout[0-9]*' $f/_rels/../_rels/$(basename $f).rels 2>/dev/null || echo "  (no layout ref found)"
done

# What theme colors are defined?
grep -A20 'a:clrScheme' /home/claude/unpacked/ppt/theme/theme1.xml | grep -E 'val=|name='
```

Read `references/template-analysis.md` for full inspection guidance: color
extraction, font detection, placeholder mapping, and capacity estimation.

---

## Step 2 — Plan Content

**Build a written slide plan before touching any XML.** This prevents
structural rework after content has been populated.

For each slide in the plan, record:
- Slide number → source template slide (e.g., "Slide 3 → use slide2.xml layout")
- Topic / section title
- Content type: stat, bullets, quote, two-column, image+text, etc.
- Approximate content length (word/character count estimate)

Rules:
- Use every distinct layout at least once if the content fits
- Never use the same layout for two consecutive content slides
- Match content type to layout (stats → callout, team → multi-column)
- For typical business presentations: 8–12 slides; pitch decks: 10–15

Read `references/content-mapping.md` for layout-to-content matching tables,
standard slide structures (business, pitch, educational, status report), and
content density rules.

---

## Step 2.5 — Measure Placeholder Capacity (Required Before Writing)

**This step must happen before editing any slide XML.**

Text overflow is the most common and most preventable visual failure. Catching
it before writing content avoids multiple fix cycles.

### Measure box dimensions

```bash
# For a specific slide, get all shape names, widths, and heights
python3 -c "
import xml.etree.ElementTree as ET
NS = {'a':'http://schemas.openxmlformats.org/drawingml/2006/main',
      'p':'http://schemas.openxmlformats.org/presentationml/2006/main'}
tree = ET.parse('/home/claude/unpacked/ppt/slides/slide2.xml')
for sp in tree.findall('.//p:sp', NS):
    name = sp.find('.//p:cNvPr', NS)
    ext  = sp.find('.//a:ext', NS)
    if name is not None and ext is not None:
        cx = int(ext.get('cx',0)); cy = int(ext.get('cy',0))
        print(f\"{name.get('name','?'):40s}  {cx/914400:.2f}\" x {cy/914400:.2f}\")
"
```

### Set a character budget

Use this table to determine how much text fits safely (≤ 80% capacity rule):

| Box height | Box width | Font size | Max lines | Chars/line | Safe budget (80%) |
|------------|-----------|-----------|-----------|------------|-------------------|
| 0.5"       | 3"        | 14pt      | 2         | 31         | ~50               |
| 0.5"       | 4"        | 12pt      | 2         | 48         | ~77               |
| 1"         | 3"        | 14pt      | 5         | 31         | ~124              |
| 1"         | 3"        | 18pt      | 4         | 24         | ~77               |
| 1"         | 4"        | 14pt      | 5         | 41         | ~164              |
| 2"         | 3"        | 14pt      | 10        | 31         | ~248              |
| 2"         | 3"        | 18pt      | 7         | 24         | ~134              |
| 2"         | 4"        | 16pt      | 9         | 34         | ~245              |
| 3"         | 4"        | 16pt      | 13        | 34         | ~354              |
| 4"         | 4"        | 16pt      | 17        | 34         | ~462              |

Formula: `capacity = (height_in / (font_pt * 1.2 / 72)) * (width_in / (font_pt * 0.5 / 72))`

**If your planned text exceeds the 80% budget, shorten it NOW before writing XML.**

### Run the automated audit

After populating slides (before packing), run the overflow audit:

```bash
python /mnt/skills/user/generating-pptx-from-template/scripts/audit_overflow.py \
  /home/claude/unpacked/ppt/slides/
```

This reports every text box as:
- 🔴 **OVERFLOW** — content already exceeds the box (must fix before delivering)
- 🟡 **AT RISK** — content at 85–100% capacity (review visually)
- And provides a recommended maximum font size for each flagged box

Fix all 🔴 items. Visually verify all 🟡 items in the QA render.

See `references/content-editing.md` → **Text Overflow — Full Fix Protocol** for
the complete decision tree covering all four fix strategies:
autofit XML, font size reduction, box expansion, and content shortening.

---

## Step 3 — Structural Edits

**Complete ALL structural operations before editing any content.**
Editing content before structure is done causes XML corruption and lost work.

Structural order:
1. Remove slides not in your plan (edit `<p:sldIdLst>` in `ppt/presentation.xml`)
2. Duplicate slides you need to reuse: `python scripts/add_slide.py unpacked/ slideN.xml`
3. Reorder slides by rearranging `<p:sldId>` entries in `presentation.xml`

After all structural changes, always clean:

```bash
python /mnt/skills/public/pptx/scripts/clean.py /home/claude/unpacked/
```

---

## Step 3.5 — Populate Content

### Editing rules

**Preferred method:** Use `str_replace` to make targeted replacements in slide XML.
This is the safest approach for small, clearly unique text blocks.

**Acceptable method:** Use Python for batch replacements across a slide when
many text values need changing and `str_replace` would require too many
individual calls. Always write the full file back atomically:

```python
with open(slide_path, 'r', encoding='utf-8') as f:
    content = f.read()
# make all replacements
content = content.replace('<a:t>OLD TEXT</a:t>', '<a:t>NEW TEXT</a:t>', 1)
with open(slide_path, 'w', encoding='utf-8') as f:
    f.write(content)
```

**Never use:** `sed`, `awk`, or regex substitutions that modify surrounding XML.
These corrupt namespace declarations and run-property attributes unpredictably.

### Critical XML rules

- Bold all titles and section headers: `b="1"` on `a:rPr`
- Never use unicode bullet characters — use `<a:buChar char="•"/>` or `<a:buAutoNum>`
- Create separate `<a:p>` elements per bullet — never concatenate into one string
- Preserve `a:pPr` blocks (line spacing, indentation, alignment) from originals
- Use theme color references (`<a:schemeClr val="accent1"/>`) not hardcoded hex values
- Remove `<a:latin typeface="..."/>` overrides to inherit theme fonts automatically
- Use XML entities for special characters: `&#x201C;` `&#x201D;` `&#x2014;` `&#x2019;`
- For text with intentional leading/trailing spaces: `<a:t xml:space="preserve"> text</a:t>`

### After populating each slide

Run the overflow audit immediately:
```bash
python /mnt/skills/user/generating-pptx-from-template/scripts/audit_overflow.py \
  /home/claude/unpacked/ppt/slides/
```

Fix all 🔴 OVERFLOW items before moving to the next slide.

### Pack the output

```bash
python /mnt/skills/public/pptx/scripts/office/pack.py \
  /home/claude/unpacked/ \
  /home/claude/output.pptx \
  --original /home/claude/template.pptx
```

---

## Step 4 — QA (Required — Do Not Skip)

### Content check

```bash
# Check for leftover placeholder text
python /mnt/skills/user/generating-pptx-from-template/scripts/check_placeholders.py \
  /home/claude/output.pptx

# Also run a manual scan for common patterns
python -m markitdown /home/claude/output.pptx | \
  grep -iE "lorem|ipsum|click to|placeholder|your title|add text|xxxx|20xx|20yy"
```

Fix any matches before continuing.

### Visual render

```bash
cd /home/claude
python /mnt/skills/public/pptx/scripts/office/soffice.py --headless --convert-to pdf output.pptx
pdftoppm -jpeg -r 150 output.pdf slide
# Inspect: slide-1.jpg, slide-2.jpg, ...
```

### Visual QA checklist — inspect every slide image for:

**Layout & brand:**
- [ ] Template brand elements preserved (colors, fonts, logos, shapes)
- [ ] Background not replaced or overridden
- [ ] Consistent margins (minimum 0.5 inch from edges)
- [ ] Layout variety across slides — not every slide the same structure

**Text overflow — check every text box:**
- [ ] No text cut off at the bottom of any box
- [ ] No stat numbers or titles clipped mid-character or mid-word
- [ ] No text overlapping an adjacent image or shape
- [ ] No small caption/footer boxes overflowing into other zones
- [ ] All bullet points fully visible (last bullet not clipped)

**Content quality:**
- [ ] No leftover placeholder text (lorem, xxxx, "Click to add", etc.)
- [ ] Readable contrast on all text against its background
- [ ] Numbers have units (%, $, Bn) — never bare numbers
- [ ] Titles are descriptive, not generic

### Fixing overflow found in the visual render

1. Match the overflowing shape in the image to its XML using `off x=` / `off y=` coordinates
2. Apply the decision tree from `references/content-editing.md` → Text Overflow section:
   - **Body/description/bullet box** → add `<a:normAutofit/>` to `<a:bodyPr>`
   - **Stat callout or title (design-critical)** → shorten the text content
   - **Spare vertical space available** → increase `cy` in `<a:ext>`
   - **No space to expand** → reduce `sz` by 100–300 per `<a:rPr>` + shorten text
3. Re-pack and re-render the affected slide
4. Confirm fix is resolved before declaring QA passed

**Do not declare success without at least one full fix-and-verify cycle.**

---

## Step 5 — Deliver

```bash
cp /home/claude/output.pptx /mnt/user-data/outputs/presentation.pptx
```

Present `/mnt/user-data/outputs/presentation.pptx` to the user with a brief
summary of what each slide covers.

---

## Multiple Templates

If the user uploads more than one `.pptx` file:

1. Analyze all templates with markitdown and thumbnails
2. Select the richest template (most layout variety, strongest brand) as primary
3. Import unique layouts from secondary templates only when needed

Read `references/multi-template.md` for the full merge procedure including
media file handling, layout reference resolution, and Content_Types.xml registration.

---

## Common Failure Modes

| Problem | Root Cause | Fix |
|---------|-----------|-----|
| Output colors changed | Hardcoded hex in slide XML | Use `<a:schemeClr>` references instead |
| Fonts changed | Hardcoded `<a:latin typeface>` override | Remove the override — let theme inherit |
| Template logos/images missing | Media files moved or deleted | Never delete `ppt/media/`; always use `--original` flag in pack.py |
| Text cut off at box bottom | Content length exceeds box height | Add `<a:normAutofit/>` to `<a:bodyPr>`, or reduce `sz`, or increase `cy` |
| Text overlaps adjacent shape | Box expanded into another shape's lane | Check `off` + `ext` positions; shrink content or reduce `cy` |
| Stat number clipped | Large `sz` + long label doesn't fit box width | Shorten label; reduce stat `sz` by 200–400; never autofit stat boxes |
| Title truncated mid-word | Title text too long for placeholder width | Rewrite title to ≤ 40 chars; reduce sz to minimum 2800 |
| Content unreadably tiny | Over-reduced `sz` to force fit | Expand `cy` if space allows; split content across two slides |
| Wrong slide duplicated | add_slide.py called with wrong source | Confirm source slide filename before duplicating |
| Structural edits lost | Content edited before structure complete | Always complete all add/remove/reorder before any text edits |
| XML corruption | Used sed/awk/regex on XML | Only use str_replace or atomic Python file writes |

---

## Reference Files

| File | Contents |
|------|----------|
| `references/template-analysis.md` | Template inspection: color extraction, font detection, placeholder mapping, capacity estimation |
| `references/content-editing.md` | XML patterns for all content types + full text overflow fix protocol |
| `references/content-mapping.md` | Layout-to-content matching, presentation structures, density rules |
| `references/multi-template.md` | Merging multiple uploaded templates |
| `scripts/audit_overflow.py` | Automated overflow detection: scans slide XML, reports at-risk boxes with recommended font sizes |
| `scripts/check_placeholders.py` | QA gate: scans output for leftover placeholder text patterns |

name: generating-pptx-from-template description: "Generates fully populated PowerPoint presentations that inherit the exact colors, fonts, layouts, and branding from a user-supplied .pptx template. Use when the user uploads or references a .pptx template file and asks to create or generate a presentation on a given topic. Triggers on: use my template, create slides from my pptx, generate a deck using my template, fill in my template, make a presentation from the attached file, build slides using my brand template, or when the user shares a .pptx file and wants a polished output matching their brand. Also use when multiple .pptx files are uploaded and the user wants a presentation combining layouts from all of them." license: MIT

Generating PPTX from Template

Produces polished PowerPoint presentations using an uploaded .pptx file as the design source of truth. The output inherits the template's exact visual identity — no generic AI aesthetics, no hardcoded colors or fonts.

Setup

Install required tools before any other step:

pip install "markitdown[pptx]" Pillow defusedxml --break-system-packages -q
which pdftoppm || apt-get install -y poppler-utils -q

All helper scripts live in two locations:

# Public pptx skill scripts (structural operations)
/mnt/skills/public/pptx/scripts/office/unpack.py
/mnt/skills/public/pptx/scripts/office/pack.py
/mnt/skills/public/pptx/scripts/office/soffice.py
/mnt/skills/public/pptx/scripts/add_slide.py
/mnt/skills/public/pptx/scripts/clean.py
/mnt/skills/public/pptx/scripts/thumbnail.py

# This skill's scripts (QA and analysis)
/mnt/skills/user/generating-pptx-from-template/scripts/audit_overflow.py
/mnt/skills/user/generating-pptx-from-template/scripts/check_placeholders.py

Workflow Overview

Step 1  →  Ingest template (markitdown + thumbnails + unpack)
Step 2  →  Plan content (layout inventory + slide map)
Step 2.5 → Measure placeholder capacity BEFORE writing any content
Step 3  →  Structural edits (add/remove/reorder slides)
Step 3.5 → Populate content (XML editing with overflow prevention)
Step 4  →  QA (content check + visual render + overflow check)
Step 5  →  Deliver

Step 1 — Ingest the Template

Locate and copy the uploaded template:

ls /mnt/user-data/uploads/*.pptx
cp /mnt/user-data/uploads/YOUR_TEMPLATE.pptx /home/claude/template.pptx

Extract all text and structure to understand existing content:

python -m markitdown /home/claude/template.pptx

Generate visual thumbnails — this is mandatory, not optional. You must see the actual layout, colors, design motifs, and photo placement before writing a single character of content:

python /mnt/skills/public/pptx/scripts/thumbnail.py \
  /home/claude/template.pptx /home/claude/thumbnails

Unpack to XML for editing:

python /mnt/skills/public/pptx/scripts/office/unpack.py \
  /home/claude/template.pptx /home/claude/unpacked/

After unpacking, run a quick structural survey:

# How many slides?
ls /home/claude/unpacked/ppt/slides/slide*.xml | grep -v _rels

# What layouts does each slide use?
for f in /home/claude/unpacked/ppt/slides/slide*.xml; do
  echo "$f:"; grep -o 'slideLayout[0-9]*' $f/_rels/../_rels/$(basename $f).rels 2>/dev/null || echo "  (no layout ref found)"
done

# What theme colors are defined?
grep -A20 'a:clrScheme' /home/claude/unpacked/ppt/theme/theme1.xml | grep -E 'val=|name='

Read references/template-analysis.md for full inspection guidance: color extraction, font detection, placeholder mapping, and capacity estimation.

Step 2 — Plan Content

Build a written slide plan before touching any XML. This prevents structural rework after content has been populated.

For each slide in the plan, record:

Slide number → source template slide (e.g., "Slide 3 → use slide2.xml layout")
Topic / section title
Content type: stat, bullets, quote, two-column, image+text, etc.
Approximate content length (word/character count estimate)

Rules:

Use every distinct layout at least once if the content fits
Never use the same layout for two consecutive content slides
Match content type to layout (stats → callout, team → multi-column)
For typical business presentations: 8–12 slides; pitch decks: 10–15

Read references/content-mapping.md for layout-to-content matching tables, standard slide structures (business, pitch, educational, status report), and content density rules.

Step 2.5 — Measure Placeholder Capacity (Required Before Writing)

This step must happen before editing any slide XML.

Text overflow is the most common and most preventable visual failure. Catching it before writing content avoids multiple fix cycles.

Measure box dimensions

# For a specific slide, get all shape names, widths, and heights
python3 -c "
import xml.etree.ElementTree as ET
NS = {'a':'http://schemas.openxmlformats.org/drawingml/2006/main',
      'p':'http://schemas.openxmlformats.org/presentationml/2006/main'}
tree = ET.parse('/home/claude/unpacked/ppt/slides/slide2.xml')
for sp in tree.findall('.//p:sp', NS):
    name = sp.find('.//p:cNvPr', NS)
    ext  = sp.find('.//a:ext', NS)
    if name is not None and ext is not None:
        cx = int(ext.get('cx',0)); cy = int(ext.get('cy',0))
        print(f\"{name.get('name','?'):40s}  {cx/914400:.2f}\" x {cy/914400:.2f}\")
"

Set a character budget

Use this table to determine how much text fits safely (≤ 80% capacity rule):

Box height	Box width	Font size	Max lines	Chars/line	Safe budget (80%)
0.5"	3"	14pt	2	31	~50
0.5"	4"	12pt	2	48	~77
1"	3"	14pt	5	31	~124
1"	3"	18pt	4	24	~77
1"	4"	14pt	5	41	~164
2"	3"	14pt	10	31	~248
2"	3"	18pt	7	24	~134
2"	4"	16pt	9	34	~245
3"	4"	16pt	13	34	~354
4"	4"	16pt	17	34	~462

Formula: capacity = (height_in / (font_pt * 1.2 / 72)) * (width_in / (font_pt * 0.5 / 72))

If your planned text exceeds the 80% budget, shorten it NOW before writing XML.

Run the automated audit

After populating slides (before packing), run the overflow audit:

python /mnt/skills/user/generating-pptx-from-template/scripts/audit_overflow.py \
  /home/claude/unpacked/ppt/slides/

This reports every text box as:

🔴 OVERFLOW — content already exceeds the box (must fix before delivering)
🟡 AT RISK — content at 85–100% capacity (review visually)
And provides a recommended maximum font size for each flagged box

Fix all 🔴 items. Visually verify all 🟡 items in the QA render.

See references/content-editing.md → Text Overflow — Full Fix Protocol for the complete decision tree covering all four fix strategies: autofit XML, font size reduction, box expansion, and content shortening.

Step 3 — Structural Edits

Complete ALL structural operations before editing any content. Editing content before structure is done causes XML corruption and lost work.

Structural order:

Remove slides not in your plan (edit <p:sldIdLst> in ppt/presentation.xml)
Duplicate slides you need to reuse: python scripts/add_slide.py unpacked/ slideN.xml
Reorder slides by rearranging <p:sldId> entries in presentation.xml

After all structural changes, always clean:

python /mnt/skills/public/pptx/scripts/clean.py /home/claude/unpacked/

Step 3.5 — Populate Content

Editing rules

Preferred method: Use str_replace to make targeted replacements in slide XML. This is the safest approach for small, clearly unique text blocks.

Acceptable method: Use Python for batch replacements across a slide when many text values need changing and str_replace would require too many individual calls. Always write the full file back atomically:

with open(slide_path, 'r', encoding='utf-8') as f:
    content = f.read()
# make all replacements
content = content.replace('<a:t>OLD TEXT</a:t>', '<a:t>NEW TEXT</a:t>', 1)
with open(slide_path, 'w', encoding='utf-8') as f:
    f.write(content)

Never use: sed, awk, or regex substitutions that modify surrounding XML. These corrupt namespace declarations and run-property attributes unpredictably.

Critical XML rules

Bold all titles and section headers: b="1" on a:rPr
Never use unicode bullet characters — use <a:buChar char="•"/> or <a:buAutoNum>
Create separate <a:p> elements per bullet — never concatenate into one string
Preserve a:pPr blocks (line spacing, indentation, alignment) from originals
Use theme color references (<a:schemeClr val="accent1"/>) not hardcoded hex values
Remove <a:latin typeface="..."/> overrides to inherit theme fonts automatically
Use XML entities for special characters: “ ” — ’
For text with intentional leading/trailing spaces: <a:t xml:space="preserve"> text</a:t>

After populating each slide

Run the overflow audit immediately:

python /mnt/skills/user/generating-pptx-from-template/scripts/audit_overflow.py \
  /home/claude/unpacked/ppt/slides/

Fix all 🔴 OVERFLOW items before moving to the next slide.

Pack the output

python /mnt/skills/public/pptx/scripts/office/pack.py \
  /home/claude/unpacked/ \
  /home/claude/output.pptx \
  --original /home/claude/template.pptx

Step 4 — QA (Required — Do Not Skip)

Content check

# Check for leftover placeholder text
python /mnt/skills/user/generating-pptx-from-template/scripts/check_placeholders.py \
  /home/claude/output.pptx

# Also run a manual scan for common patterns
python -m markitdown /home/claude/output.pptx | \
  grep -iE "lorem|ipsum|click to|placeholder|your title|add text|xxxx|20xx|20yy"

Fix any matches before continuing.

Visual render

cd /home/claude
python /mnt/skills/public/pptx/scripts/office/soffice.py --headless --convert-to pdf output.pptx
pdftoppm -jpeg -r 150 output.pdf slide
# Inspect: slide-1.jpg, slide-2.jpg, ...

Visual QA checklist — inspect every slide image for:

Layout & brand:

Template brand elements preserved (colors, fonts, logos, shapes)
Background not replaced or overridden
Consistent margins (minimum 0.5 inch from edges)
Layout variety across slides — not every slide the same structure

Text overflow — check every text box:

No text cut off at the bottom of any box
No stat numbers or titles clipped mid-character or mid-word
No text overlapping an adjacent image or shape
No small caption/footer boxes overflowing into other zones
All bullet points fully visible (last bullet not clipped)

Content quality:

No leftover placeholder text (lorem, xxxx, "Click to add", etc.)
Readable contrast on all text against its background
Numbers have units (%, $, Bn) — never bare numbers
Titles are descriptive, not generic

Fixing overflow found in the visual render

Match the overflowing shape in the image to its XML using off x= / off y= coordinates
Apply the decision tree from references/content-editing.md → Text Overflow section:
- Body/description/bullet box → add <a:normAutofit/> to <a:bodyPr>
- Stat callout or title (design-critical) → shorten the text content
- Spare vertical space available → increase cy in <a:ext>
- No space to expand → reduce sz by 100–300 per <a:rPr> + shorten text
Re-pack and re-render the affected slide
Confirm fix is resolved before declaring QA passed

Do not declare success without at least one full fix-and-verify cycle.

Step 5 — Deliver

cp /home/claude/output.pptx /mnt/user-data/outputs/presentation.pptx

Present /mnt/user-data/outputs/presentation.pptx to the user with a brief summary of what each slide covers.

Multiple Templates

If the user uploads more than one .pptx file:

Analyze all templates with markitdown and thumbnails
Select the richest template (most layout variety, strongest brand) as primary
Import unique layouts from secondary templates only when needed

Read references/multi-template.md for the full merge procedure including media file handling, layout reference resolution, and Content_Types.xml registration.

Common Failure Modes

Problem	Root Cause	Fix
Output colors changed	Hardcoded hex in slide XML	Use `<a:schemeClr>` references instead
Fonts changed	Hardcoded `<a:latin typeface>` override	Remove the override — let theme inherit
Template logos/images missing	Media files moved or deleted	Never delete `ppt/media/`; always use `--original` flag in pack.py
Text cut off at box bottom	Content length exceeds box height	Add `<a:normAutofit/>` to `<a:bodyPr>`, or reduce `sz`, or increase `cy`
Text overlaps adjacent shape	Box expanded into another shape's lane	Check `off` + `ext` positions; shrink content or reduce `cy`
Stat number clipped	Large `sz` + long label doesn't fit box width	Shorten label; reduce stat `sz` by 200–400; never autofit stat boxes
Title truncated mid-word	Title text too long for placeholder width	Rewrite title to ≤ 40 chars; reduce sz to minimum 2800
Content unreadably tiny	Over-reduced `sz` to force fit	Expand `cy` if space allows; split content across two slides
Wrong slide duplicated	add_slide.py called with wrong source	Confirm source slide filename before duplicating
Structural edits lost	Content edited before structure complete	Always complete all add/remove/reorder before any text edits
XML corruption	Used sed/awk/regex on XML	Only use str_replace or atomic Python file writes

Reference Files

File	Contents
`references/template-analysis.md`	Template inspection: color extraction, font detection, placeholder mapping, capacity estimation
`references/content-editing.md`	XML patterns for all content types + full text overflow fix protocol
`references/content-mapping.md`	Layout-to-content matching, presentation structures, density rules
`references/multi-template.md`	Merging multiple uploaded templates
`scripts/audit_overflow.py`	Automated overflow detection: scans slide XML, reports at-risk boxes with recommended font sizes
`scripts/check_placeholders.py`	QA gate: scans output for leftover placeholder text patterns