Fixing Google Slides PDF to HTML Export Issues with the Slides API

by

Summary

The workflow of exporting a Google Slides deck to PDF, feeding it to Claude for HTML conversion, and then updating charts is fragile. Claude’s hallucinations during PDF‑to‑HTML reconstruction cause misplaced elements, incorrect colors, and missing graphics, making the process unmaintainable across decks.

Root Cause

  • Claude treats the PDF as unstructured input and attempts to infer layout, which is lossy.
  • PDF does not retain the semantic information (slide IDs, chart objects) needed for deterministic updates.
  • The prompt chain lacks a stable contract between the source deck and the transformation step, so any variation in layout breaks the pipeline.

Why This Happens in Real Systems

  • Many LLM‑based automation pipelines rely on format conversion (PDF → HTML/Markdown) as an intermediate step, assuming the model can perfectly reconstruct the original structure.
  • Real‑world documents contain complex styling, embedded images, and vector graphics that are not faithfully represented in plain text.
  • The model’s hallucination safety mechanisms prioritize producing plausible output over strict fidelity to the source when the input is ambiguous.

Real-World Impact

  • Repeated prompt engineering consumes engineering time and delays monthly reporting.
  • Inconsistent decks erode leadership confidence in data delivery.
  • Higher error rate: misplaced charts can lead to misinterpretation of key metrics.
  • Scalability bottleneck: each new deck template requires a custom prompt fix.

Example or Code (if necessary and relevant)

How Senior Engineers Fix It

  • Skip PDF → HTML: Use the Google Slides API to read/write slide objects directly.
  • Store a template deck in Slides and programmatically replace chart data sources via the Slides API (or via linked BigQuery data sources).
  • Leverage chart placeholders with consistent object IDs so a script can locate and update them without layout inference.
  • If HTML is required, export slides as Google Slides JSON (slides.get) rather than PDF; this retains structural metadata.
  • Automate the pipeline with a CI/CD job that:
    1. Pulls the template deck ID.
    2. Runs the BigQuery query, saves results to a CSV/JSON.
    3. Calls presentations.batchUpdate to replace chart data ranges.
    4. Publishes the updated deck.

Why Juniors Miss It

  • They often focus on LLM output rather than the underlying data contract, assuming the model can “understand” the PDF layout.
  • Lack of familiarity with Google Slides API leads to reliance on brittle conversion hacks.
  • Tendency to treat prompts as the only solution instead of establishing a deterministic, programmatic interface.

Leave a Comment