Presentation Deck Analysis

Contra Labs needed AI-generated presentation decks evaluated by someone who could judge them the way a real client would. Sixty-three decks across twenty-one distinct creative briefs, each scored and critiqued against five dimensions: preference, effectiveness, content quality, narrative quality, and visual design.

Fixed scope, fixed deadline, delivered on schedule.

Dimension-based scoring Each output assessed only on the dimensions genuinely relevant to it, rather than forcing every deck through an identical rubric. A pitch deck and an internal update fail in different ways and should not be measured the same way.

Source verification Every output checked line by line against the brief that produced it. Required content, structure, tone, word counts, named requirements, stated figures. Most evaluation stops at whether something looks good. The more useful question is whether it did what was actually asked.

Written critique Strengths and weaknesses for every output, cited specifically: slide numbers, exact phrases, precise design terminology. Written to be read by both technical and non-technical stakeholders, so the design language is there when it earns its place and explained when it does not.

The recurring failure patterns turned out to be more valuable than the scores. Across sixty-three decks the same tells kept surfacing:

Fabricated statistics presented as fact, formatted confidently enough to survive a casual read

Internally inconsistent naming, the same product called two different things inside one deck

Formulaic phrasing patterns, including a distinctive over-reliance on em dash constructions

Unrequested scope invention, slides answering questions nobody asked

Structural padding that fills a deck without advancing its argument

Rendering errors that stay invisible unless you compare the output against the original prompt

That list is the real deliverable. Scores tell a team where they are. Failure patterns tell them what to fix.

Judging a deck means judging typography, alignment, colour relationships and information hierarchy, and then judging whether any of it actually serves the argument. Twenty years of doing that work by hand is the difference between "slide four feels cluttered" and naming exactly which hierarchy decision broke and why.

Catching a fabricated statistic needs rigour. Catching a deck that is well built and still not persuasive needs a designer.

This engagement is covered by an NDA, so no deck content, brief material or client output appears here. The methodology above is my own and is described in general terms only.