Extract a web-hosted deck slide by slide

Why use this

1. A presentation embedded in a flipbook may be readable on screen yet difficult to search or compare. Charts, footnotes and slide numbering are especially vulnerable when text extraction treats the viewer as one long page rather than a sequence of distinct slides.

2. This job could create a navigable research copy whose content stays tied to the original slide. Flagging uncertain figure labels and unreadable regions would make the result more useful for careful review than a seamless transcript that conceals extraction gaps.

Why an agent / simpler alternative

Useful judgment

Interpret page structure and keep text, figures and footnotes associated with the correct slide.

Simpler alternative

Download an authorized original PDF when the publisher provides one.

Unsuitable when

The viewer restricts access or the task requires reproducing material beyond permitted use.

Proposed inputs

1. Name

   Deck reference

   Owner

   Authorized reader

   Minimum scope

   Accessible viewer link or supplied page images and known slide count

   Quality check

   Confirm deck title and version before extraction.

2. Name

   Extraction specification

   Owner

   Research owner

   Minimum scope

   Required text, figure labels, reading order and citation format

   Quality check

   Separate verbatim recovery from optional summary requests.

Access and setup — proposed or unknown

Proposed minimum

1. Read authorized slide images and save a private numbered extraction; no access bypass or republication.

Compatibility unknowns

1. Unknown: viewer navigation, image resolution, OCR quality and export compatibility.

External actions requiring approval

1. Downloading restricted assets or distributing the recovered deck content.

Expected deliverable and review checks

Proposed artifact

Slide-indexed text and figure inventory with uncertain readings and original-page references.

Critical checks

1. Every slide is accounted for, including blank or unreadable pages.

2. Chart axes, units, legends and footnotes stay attached to the right figure.

Missing output evidence

This overlay adds no new output inspection or execution validation. Preserve the canonical implementation evidence and its outcome limits.

Failure signals and proposed responses

1. Signal

   The viewer skips a page during navigation.

   Why it matters

   Slide numbering can shift and misattribute all later content.

   Proposed response

   Reconcile the page index against visible counters or a supplied manifest before continuing.

2. Signal

   A chart axis label is read as body text.

   Why it matters

   A figure’s meaning may be distorted despite accurate individual words.

   Proposed response

   Retain its spatial association and flag the label for visual review.

Effort and cost — unknown

Setup work

1. Define numbering rules and a figure-description convention.

Ongoing review

1. Review OCR uncertainty and reconcile missing or repeated pages.

Unknowns

1. Unknown: image quality, extraction charges and permitted reuse scope.

First check — proposed, unexecuted

Status

proposed_unexecuted

Synthetic fixture

Synthetic slide images include a chart with a footnote, a blank divider and two slides sharing the same heading.

Acceptance checks

1. The divider remains in the slide count.

2. Repeated headings do not merge slides, and chart labels remain grouped with the figure.

Stop conditions

1. Unreadable content would need to be guessed or restricted access bypassed.

Limits

1. Editorial extraction checks preserve provenance; they do not authorize copying or establish perfect visual recovery.