Converting Drafts to Slide Decks

Given a draft document where each section/paragraph corresponds to one slide, produce a PPTX where:

• Every slide has a clear title, 3-5 content bullets (or a visual + short bullets), and no slide feels sparse or overstuffed.
• Wording is tightened, jargon-free where possible, and consistent in tone/tense across all slides.
• Visual hierarchy (title, subtitle, body) is respected — no walls of text, no single-line slides.

Minimal input needed from user:

1. The draft document (with slide-by-slide content already written).
2. Target slide count (or "use as many as the draft implies").
3. Preferred deck style: corporate, academic/research, or pitch-style. Default to academic/research if the topic involves papers/screening/LLM pipelines, since that's the common use case here.

Progress:

• Step 1: Parse the draft into slide units
• Step 2: Normalize each slide's content to a target density
• Step 3: Rewrite text to remove AI-sounding patterns
• Step 4: Balance visuals vs. text per slide
• Step 5: Apply consistent structure/template across the deck
• Step 6: Final pass — check for empty, overloaded, or inconsistent slides

Step 1: Parse the draft into slide units

• Split the document wherever the user marked a new slide (headings, "Slide 1:", horizontal rules, etc.).
• If no explicit markers exist, infer slide breaks from headings or topic shifts — one slide per major idea, not per paragraph.
• Extract for each unit: intended title, key points, any data/numbers, any figures/diagrams mentioned.

Step 2: Normalize each slide's content to a target density

• Target per slide: 1 title + 3-5 bullets, each bullet 6-14 words.
• If a slide's draft content is thin (e.g., only one sentence), do NOT pad with filler adjectives. Instead:
  • Pull in relevant context from adjacent paragraphs in the draft (methodology detail, example, or number).
  • Add a supporting sub-point that explains "why it matters" or "how it works," grounded in the draft's own content.
  • If truly nothing to add, merge it with the next/previous slide rather than leaving it empty.
• If a slide's draft content is too dense (a full paragraph), split into:
  • Main slide with top-level bullets.
  • Optional secondary slide ("Slide 4a") for the details, only if the draft signals it's important enough.

Step 3: Rewrite text to remove AI-sounding patterns

Avoid these tells:

• Generic openers: "In today's world," "It is important to note that," "Furthermore," used repeatedly.
• Overuse of triplets ("fast, efficient, and reliable") on every slide.
• Perfectly symmetric bullet lengths (real human decks vary length).
• Overly formal connectors ("Moreover," "Additionally," "In conclusion") on every single slide.

Replace with:

• Direct, concrete phrasing tied to the actual project (e.g., "Screened 1,200 papers by title/abstract using GPT-4o" not "Leveraged advanced AI to process a large corpus of documents").
• Domain-specific terms from the user's own draft — reuse their vocabulary, don't over-synonymize.
• Varied sentence starts and bullet lengths within the same slide.

Step 4: Balance visuals vs. text per slide

• If the draft mentions a process/pipeline (e.g., paper screening flow), convert to a simple flow diagram description (Input → Screening by title/abstract → Filtered set → Output) instead of a text list — reduces text density and looks more professional.
• If the draft mentions numbers/results, suggest a simple chart or stat callout instead of a sentence.
• Rule of thumb: any slide with >5 bullets should have at least one item converted to a visual element or split out.

Step 5: Apply consistent structure across the deck

Use one of these fixed patterns per slide, chosen based on content type, and keep it consistent for the same content type throughout:

• Section/Agenda slide: title + numbered list of sections.
• Concept slide: title + 3-4 bullets + optional one visual.
• Process slide: title + diagram/flow + 1-2 supporting bullets.
• Data/results slide: title + chart/table + 1-2 line takeaway.
• Closing slide: summary of 3 key takeaways + next steps.

Keep consistent across the whole deck:

• Same title casing style (e.g., Title Case throughout).
• Same bullet punctuation (either all end with periods or none do).
• Same verb tense within similar slide types (e.g., all methodology slides in past tense: "Collected," "Filtered," "Screened").

Step 6: Final pass checklist

• No slide has fewer than 3 lines of content or a single giant image with no text.
• No slide has more than ~6 bullets or a dense paragraph.
• Terminology matches across slides (e.g., don't call it "LLM model" on one slide and "AI system" on another — pick one).
• Every process/methodology slide follows the same visual pattern.
• Read the deck top-to-bottom as a story: Problem → Method → Process → Results → Conclusion.

Example 1: Input (draft excerpt):

Slide 3: Methodology. We collect papers from various databases based on topic. Then we use LLM to screen title and abstract.

Output (slide content):

• Title: Methodology: Automated Paper Screening
• Bullets:
  • Collected papers from academic databases based on the target research topic
  • Used an LLM to screen titles and abstracts against inclusion criteria
  • Filtered out irrelevant papers before full-text review
  • Reduced manual screening time from days to hours
• Suggested visual: 3-step flow diagram (Input Papers → LLM Screening → Filtered Set)

Example 2: Input (draft excerpt, too dense):

Slide 5: The system takes a large set of input papers on a given topic, uses an LLM to read the title and abstract of each one, decides if it matches the topic and inclusion criteria, and then returns only the papers that passed this filter to the user, who can then review them manually or proceed to full-text screening depending on project needs.

Output: split into two slides.

• Slide 5 — How the Screening Pipeline Works: 4 bullets (Input → LLM reads title/abstract → Relevance decision → Filtered output), with flow diagram.
• Slide 6 — After Screening: 2 bullets (manual review option, full-text screening option) + short note on flexibility for different project needs.