Flight Check — Pre-Flight Review for Decks

A final-pass review for slide decks before they go in front of a real audience. Built by Analyst Academy.

This is a pre-flight check, run after the deck is built and before it taxis to the runway. The job is to catch the things that quietly destroy credibility — AI-generated phrasing the audience will spot, unsourced numbers, weak headlines that don't carry the message, and the small mistakes that signal "draft." Clear it, or ground it.

It is not a drafting tool, not a redesign tool, and not a substitute for the pilot's own walkaround.

When to run this

Run this skill when the user:

• Shares a .pptx, .pdf, or pasted slides and asks for a review
• Says things like "is this ready," "check this deck," "final pass," "anything I missed"
• Mentions AI-tells, sourcing, citations, polish, sanity check, or finalizing
• Asks whether their slides "sound like AI"

Don't run it when the user is still drafting (asking for new content, structure, or ideas). This checklist is for the near-final deck — for earlier stages, help them build first.

Flight procedure

Step 1. Identify what's in the cargo bay. Before reviewing, note the input type. This determines what you can and can't check:

• Full .pptx: Slide text, layout, charts, footers, speaker notes (if accessible) — all checkable.
• PDF export: Visible slides only. Speaker notes, comments, and edit-layer artifacts will not be visible.
• Screenshots or partial slides: Only what is visually present. Full-deck consistency checks (cross-slide number reconciliation, structural arc) are limited.
• Pasted slide text: Wording, structure, action titles, specificity, source presence — yes. Visual polish, layout, chart honesty — no.

State the input type and limitations at the top of the report. Never claim a check passed if it wasn't actually performed.

Step 2. Walk the full deck. Don't sample. The whole point of a pre-flight check is catching the one slide that breaks the deck's credibility — the one with the fabricated stat, the topic-label headline, the leftover [TBD].

Step 3. Run all six diagnostic systems below. Note specific findings as you go — quote exact text, name the exact slide number and title.

Step 4. File the report using the format at the bottom of this file. Lead with the most critical issues, not slide order. The pilot needs to know what could ground the flight, not what's on slide 1.

Step 5. Resist the urge to be diplomatic. A vague review ("consider strengthening some headlines") is useless. A pointed one ("Slide 7 title 'Market Overview' is a topic label, not a finding — what's the message?") is what makes this checklist worth running.

Evidence standard

Every reported issue must be locatable. Each finding must include at least one of:

• Slide number (and title if available)
• Exact quoted text from the slide
• A precise visual reference ("the chart on slide 12," "the second bullet on slide 4")

If you can't point to where an issue is, don't include it in the report. Generic comments like "some titles feel weak" are not findings — they're impressions.

The six diagnostic systems

1. The AI authorship check

The reader should not be able to tell whether a human or an AI wrote this deck. Audiences — especially executive ones — spot AI tells fast, and once they do, they discount everything that follows.

The principle: flag patterns of generic, inflated, or portable language. Do not flag individual words mechanically. "Leverage" appears in "leverage ratio" (legitimate). "Drive" appears in "drive revenue" (fine). The issue is when the language is repetitive, evidence-free, and could be lifted to any other deck unchanged.

Patterns that signal AI authorship:

Language tells

• Triplet adjectives: "comprehensive, robust, and scalable" / "innovative, agile, and customer-centric"
• Empty verbs with no specific object: "leverage capabilities," "unlock potential," "drive value," "optimize outcomes," "streamline operations"
• Stock openers: "In today's fast-paced world," "In an increasingly complex landscape," "It's worth noting that"
• Hedging without reason: "may potentially," "could possibly," "various," "numerous," "myriad"
• "Not just X — it's Y" or "It's not about X, it's about Y" framings, especially repeated
• Em-dashes used as connective tissue on every slide
• Sentences that sound formal but say nothing concrete: "We will deliver value through strategic alignment of capabilities"

Structural tells

• Every framework is exactly three steps, three pillars, three shifts
• Every breakdown is a 2x2 or four quadrants, even when the underlying logic isn't symmetric
• Bulleted lists where every bullet starts with the same verb form ("Identify... Analyze... Recommend...")
• Filler slides labeled "Overview," "Context," "Background" with no specific finding
• A conclusion slide that just restates the agenda

The replacement test. Pick three content slides at random. If you swap the named company, industry, or topic for any other, does the slide still make sense? If yes, the slide is filler — real analytical content shouldn't be portable across contexts.

False-positive guard. A single instance of an em-dash or one use of "leverage" is not an AI tell. Flag the pattern, not the word. If a deck has one stock phrase across 30 slides, that's a polish issue; if it has them on every slide, that's an authorship issue.

2. The source check

Every quantitative claim, every external fact, every "research shows" needs a verifiable source. Missing or fabricated sources are how a deck loses a sophisticated room — and AI-assisted decks fabricate plausible-sounding citations constantly.

Run this check in three layers, and be explicit about which layer you actually completed.

Layer 1 — Source presence. Is there a source line for every external number, external claim, and quoted study? This is the floor. Missing citations get flagged regardless of anything else.

Layer 2 — Source plausibility. Does the citation look real? Specific report names from real organizations are higher confidence ("McKinsey Global Institute, The State of Enterprise AI, 2024"). Vague citations are lower ("McKinsey, 2024" or "industry research"). AI-fabricated sources often cite real organizations but invented report titles.

Layer 3 — Source verification. Does the cited source actually exist, and does the number match what the source says? This is the only layer where you can genuinely confirm a citation is real.

Important: Layer 3 requires the ability to search the web. If you have web search available, attempt to verify a sample of the most consequential citations (top-of-deck stats, market sizing figures, recommendation-supporting numbers). If web search is not available or fails, you cannot complete Layer 3 — say so plainly.

Never claim a source was verified unless you actually verified it. State which layer of the check you completed:

• "Source check completed at Layer 1 (presence only — verification not attempted)."
• "Source check completed at Layer 2 (presence + plausibility review; verification not possible without web access)."
• "Source check completed at Layer 3 (presence + plausibility + spot verification of [list of citations] via web search)."

Common failure modes to flag specifically:

• "Studies have shown..." without naming the study
• "Industry research suggests..." with no citation
• A number on a chart with no source line
• The same source cited for every claim (suggests one paper is being stretched)
• Sources that exist but appear unread — phrasing in the deck doesn't match the source's actual language
• Specific named reports from real organizations — these are the highest-risk for AI fabrication and the ones most worth verifying

3. The action title check

(Analyst Academy core principle.) On content slides, every title should state the message of the