PPTX Grounding

PPTX Grounding

This skill handles PowerPoint (.pptx) inputs in the any-input -> grounding -> downstream research / summary / report pipeline.

Its job is not to write a polished final report. Its job is to produce a stable, structured, reusable evidence bundle for a deck. After that, the agent must read the bundle and write a real grounded.md.

When to use this skill

Use this skill for common PowerPoint decks such as:

• project updates
• proposal decks
• experiment / result review decks
• internal presentations
• lecture / meeting decks
• any .pptx where the main evidence lives in slide titles, bullets, text boxes, images, tables, charts, and speaker notes

Use this skill when slide order, per-slide structure, and per-slide assets matter.

Relationship to other grounding skills

This skill is not the same as document-grounding:

• document-grounding is primarily linear-document oriented
• pptx-grounding is slide-oriented
• slide order, slide-level structure, and per-slide assets matter here

This skill is also not table-grounding:

• it may export tables and charts found inside a deck
• but it treats them as deck evidence assets, not as standalone spreadsheet analysis tasks

Supported extraction targets

This skill should extract and organize the following when present:

• deck-level metadata
• slide-by-slide structure
• slide title extraction
• bullet / paragraph extraction
• other text-box extraction
• speaker notes extraction
• embedded image export
• native PowerPoint table export
• native PowerPoint chart export (lightweight best-effort extraction)
• AssetRef insertion into extracted.md
• bundle generation for downstream agent grounding

Out of scope

This skill does not aim to handle the following:

• video / audio extraction
• animation / transition analysis
• OCR on images
• SmartArt semantic reconstruction
• precise visual rendering / slide preview generation
• embedded file deep extraction
• comment / revision workflows
• automatic polished final report generation

Required Workflow

When using this skill, you must follow this workflow:

1. First run the existing script entrypoint:

   bash .cursor/skills/pptx-grounding/scripts/run.sh <input_pptx> <output_root>
   

2. The script generates an evidence bundle only. It does not generate grounded.md.

3. After the bundle is generated, read:

   • extracted.md
   • extracted_meta.json
   • slide_index.json
   • asset_index.json

4. If AssetRef blocks are present in extracted.md, inspect the referenced assets before writing grounded.md. This includes, when present:

   • assets/images/*
   • assets/tables/*
   • assets/charts/*

5. Then write a real grounded.md into the same bundle directory.

6. Do not create a placeholder grounded.md. If grounded.md exists, it must be a real grounding note written after inspecting the generated bundle.

7. The task is not complete if only the extraction bundle exists.

8. The task is also not complete if grounded.md is written without first generating and using the bundle evidence.

Input

bash .cursor/skills/pptx-grounding/scripts/run.sh <input_pptx> <output_root>

Arguments

• input_pptx: path to a .pptx file
• output_root: parent directory under which the bundle should be created

Example:

bash .cursor/skills/pptx-grounding/scripts/run.sh \
  /path/to/deck.pptx \
  data/grounded_notes

This will create something like:

<data/grounded_notes>/<ground_id>/

Output bundle contract

The script should create:

<pptx_bundle>/
├─ extracted.md
├─ extracted_meta.json
├─ slide_index.json
├─ asset_index.json
├─ assets/
│  ├─ images/
│  ├─ tables/
│  └─ charts/
└─ grounded.md   # written later by the agent, not by the script

File roles

extracted.md

Human-readable entrypoint for the agent.

It should include:

• deck overview
• slide-by-slide extraction
• title
• bullet blocks
• other text blocks
• notes text
• AssetRef blocks for images / tables / charts

extracted_meta.json

Global deck metadata, such as:

• source file
• source type
• deck id
• slide count
• parser name
• note / image / table / chart counts
• extraction timestamp
• extraction scope notes

slide_index.json

Machine-readable slide-level index.

Each slide record should include:

• slide number
• slide id
• layout name (if available)
• title
• bullet blocks
• text blocks
• notes text
• assets
• shape statistics

asset_index.json

Global asset registry for all exported assets.

Each asset record should include:

• asset id
• asset type
• slide number
• source shape id / name
• relative path
• lightweight metadata
• extraction status

Extraction targets

1. Slide title

Extract separately when possible.

2. Bullet / paragraph text

Preserve paragraph order and indentation level where possible.

3. Other text boxes

Extract non-title text separately from bullet blocks.

4. Speaker notes

Extract notes text when present.

5. Images

Export embedded images into assets/images/ and register them in asset_index.json.

6. Tables

For native PowerPoint tables:

• export markdown view
• export csv view
• register as table assets

7. Charts

For native PowerPoint charts, do lightweight best-effort export:

• chart metadata JSON
• flattened CSV if possible
• register as chart assets

If chart extraction is partial, record that clearly in metadata rather than pretending extraction succeeded fully.

grounded.md schema (recommended)

After reading the bundle, the agent should write a stable grounding note such as:

# PPTX Grounding

## 1. Main Topic / Purpose

## 2. Deck Structure / Narrative Flow

## 3. Main Points

## 4. Important Evidence and Assets

## 5. Speaker Notes Signals (only if present and materially useful)

## 6. Gaps / Risks / Ambiguities

## 7. Suggested Next Steps

## 8. Search Keywords

Writing guidance

grounded.md should be a stable intermediate grounding note, not a polished final report.

It should:

• summarize what the deck is mainly about
• preserve the deck's narrative flow at a useful level
• call out the most important slides and extracted evidence that materially affects interpretation
• mention notes, tables, charts, or other asset types only when they are actually present and useful
• record real ambiguity, evidence gaps, or extraction issues that matter for downstream use
• support downstream research / summary / report steps

It should not:

• pretend the bundle was fully interpreted without inspecting assets
• invent conclusions unsupported by the deck
• treat the simple absence of an asset type as a risk or extraction failure
• add a "Missing Assets" section just because a deck has no charts, no tables, or no speaker notes
• replace the downstream final report stage

Important rule on absent asset types

Do not write that charts / tables / images / speaker notes are "missing" unless there is positive evidence that the deck contains them and extraction failed.

If an asset type is simply not present in the PPTX, do one of the following:

• omit it entirely from grounded.md, or
• mention it neutrally only if necessary for clarity

Examples of acceptable neutral wording:

• No speaker notes were detected in this deck.
• No native PowerPoint charts were detected in this deck.

Do not present these as risks, failures, or missing assets by default.

Failure handling

If extraction is partial or some assets fail to export:

• still write the bundle files that were successfully produced
• record failures explicitly in asset_index.json or metadata
• do not fabricate successful extraction
• do not skip bundle generation silently

If a deck is too complex for full extraction, produce the best available bundle and make the limitations explicit.