The Human Pass: How to Edit AI-Generated Slides
An AI-generated deck is not a finished deck; it's a confident first draft with your name on it. The people who get good results from AI slides spend 20–30 minutes on a deliberate edit pass. The people who get embarrassed skip it, because draft one looks finished — complete sentences, consistent formatting, a conclusion slide — and finished-looking is exactly the trap. The problems aren't typos. They're structural: the deck covers a topic instead of making an argument, every example is generic, and one statistic that came from nowhere is waiting for the sharpest person in the room.
The pass below is ordered by leverage: fixing sentences before fixing the story is wasted effort, because half the slides won't survive anyway. It takes 25 minutes for a 10-slide deck, and it has a defined endpoint so it doesn't become an hour of rearranging.
Pass zero: read the titles, and only the titles
Before touching any slide, read only the slide titles in order, out loud. Titles are the deck's skeleton, and the skeleton tells you which kind of deck you have. AI organizes decks by topic: Background, Overview, Details, Conclusion. Rooms are moved by argument: problem, stakes, options, recommendation, ask. A topic deck is a table of contents read aloud; an argument deck is the 60-second version of your talk. If the titles don't read as that 60-second version, no amount of sentence-level editing will save the deck — fix order and merges until they do.
Two reorderings that solve most structural problems. For executive audiences, move the recommendation to slide 1 or 2 and let the rest of the deck become the evidence you call up under questions; the "build up to the conclusion" structure spends the room's attention before the point arrives. For technical audiences, keep the build-up but cut everything that isn't a finding or a decision — engineers read the deck as claims to verify, and background slides read as claims without evidence.
Point 1: Story — merge until each slide earns its place
With the skeleton sound, apply the merge rules. Two slides making the same point become one; the merged version is always stronger because it has to choose its best evidence. A slide with no point — the "Introduction", the "Overview of the landscape" — gets deleted or absorbed into a neighbor as one bullet. After each merge, re-run the title test: the argument should tighten, never wobble.
The AI's specific storytelling tells, worth hunting: an agenda slide (delete; the title test replaces it), a "Background" slide that's really three bullets of context (absorb), a "Next Steps" slide that duplicates the recommendation (keep one, and it's the recommendation), and pairs like "Current State" / "Challenges" that usually share a spine. Merging feels like losing work; it's actually the deck gaining weight.
Point 2: Specificity — hunt the generic phrases
AI writes in a dialect of safe generality, and the tells are consistent: "a leading company in the industry", "users may struggle with", "significant improvements in efficiency", "various stakeholders across the organization", "in today's fast-paced market". Each one is a claim that survived without evidence. The pass: search the deck for them, then apply the rule — replace with your real example, or delete the claim. There is no third option, because "sort of specific" reads as vague with extra steps.
"A leading company in the industry" becomes "Customer X cut onboarding from 11 days to 4" or dies. "Users may struggle" becomes the actual support ticket category, or dies. If you don't have the real example, the claim wasn't load-bearing — which is useful to learn before the room asks for it. Ten minutes of replacements turns a template-scented deck into one that could only have come from your team; this is the highest-value ten minutes in the whole pass.
Point 3: Deletion — the deck gets shorter and better
Agenda slide, "Introduction", the triple summary at the end: gone. One recap slide is allowed if the deck runs past 15 slides, and its content is the ask, not a re-listing. Deletion isn't vandalism — it's where the remaining slides get their weight. A 10-slide deck with 10 arguments reads as decisive; the same deck with 6 arguments and 4 filler slides reads as padded, and rooms discount everything in it.
The counterintuitive part: after a deletion pass, the deck is easier to present. Fewer slides means fewer transitions to fumble and fewer places for questions to ambush you. If a deleted slide contained detail someone might ask about, it moves to the appendix, not the trash — the appendix is where "what about X?" goes to die politely.
Point 4: Verification — the non-negotiable one
AI invents statistics fluently, and it doesn't flag them. One unsourced number in front of the wrong audience costs more credibility than the whole deck saved you, so this point has a procedure. First, inventory: ask the AI "List every factual claim and statistic in this deck as a checklist, with the slide it appears on." Second, triage: the five claims that carry the argument are load-bearing and get verified against a source you trust; the rest are decorative. Third, act: a decorative claim gets softened ("early results suggest…") or cut; a load-bearing claim you can't verify gets replaced with [NEED DATA] and filled by you before the deck moves.
The standard is the source-on-the-spot rule: never let a number reach a room that you couldn't source if someone asks "where's that from?" while you're standing there. "The AI said so" is not an answer you want to give out loud. The document method shows this check applied to extracted numbers; the principle is the same everywhere.
Point 5: Read-aloud — the pass that catches what eyes miss
Read the deck out loud once, at speaking pace, notes and all. Three things surface that silent reading never catches. Slides where you stumble are slides written in a register you don't speak — rewrite them in the words you actually said, which are always clearer than the AI's. Slides that take ninety seconds to read aloud are two slides or half a slide, and you now know which. And sentences that sound impressive but mean nothing — "enables seamless cross-functional alignment" — become obvious the moment your mouth has to say them to a room you respect.
Rule of thumb for length while reading: roughly 60 words of slide-plus-notes per minute of speaking. A 15-minute slot needs about 900 words across the deck; if the read-aloud takes 25 minutes, the deck is 40% too long, and you already know from Point 3 what to do about that.
The "so what?" test, slide by slide
For each slide, answer one question: if a listener saw only this slide, what should they conclude? If there's no answer, the slide is decoration — merge or delete. If the answer exists but the title doesn't state it, the title is a label wearing a suit. The test converts label titles into takeaway titles, and the difference is the difference between a deck people read and a deck people absorb:
| Label title (fails) | Takeaway title (passes) |
|---|---|
| Churn analysis | Churn doubled after the pricing change — in month 2, not month 1 |
| Competitor overview | Only one competitor bundles support; it's winning our renewals |
| Q3 marketing results | Q3 leads up 22%, but cost per lead up 40% — efficiency broke, not volume |
| Team structure | Two teams, one backlog: the bottleneck is onboarding, not capacity |
| Security review | We pass 14 of 15 controls; the gap is fixable in one sprint |
Run the test during pass zero, not at the end — takeaway titles change how the slides underneath get built, and retrofitting them onto a finished deck doubles the work.
The voice pass: making it sound like you
Decks have accents, and AI has a different one than yours. If the deck will be presented by you, it should sound like you, because the audience hears your voice and reads the slides — a mismatch registers as inauthenticity nobody can name. Two cheap fixes. First, paste a paragraph you actually wrote (an email, a doc) and ask the AI: "Rewrite the slide text and notes to match this writing style." Second, maintain the banned-phrase list from the formula article — synergy, seamless, leverage, "in today's world" — and search for them by hand, because they survive prompts and die to Ctrl+F.
Two more voice tells: the AI writes longer sentences than you speak, so split anything past twenty words; and it hedges ("may", "potentially", "could potentially") where a human presenting takes a position. Hedged slide text reads as uncertain authorship. Pick the verb, state the claim, put the uncertainty in the notes where it belongs.
Cutting from 14 slides to 10
When the draft arrives over length, the merge candidates are predictable. "Introduction" and "Overview" are one slide. Two results slides from different angles are usually one. "Key Learnings" and "Next Steps" and "Conclusion" are one slide, and it's called the ask. "Methodology" and "Data Sources" go to the appendix together. And any slide whose content is one bullet plus decoration gets absorbed into the slide that motivated it. Four merges, ten slides, no information lost — the detail is in the notes, where you'll actually use it.
What never gets cut in the pass: the slide that states the ask, the slide with the strongest number, and the slide that answers the audience's hardest known question. Everything else is negotiable; those three are the deck.
Editing the notes, not just the slides
The pass covers slides; notes need their own two minutes, and they fail differently. AI notes run longer than you'll actually speak, hedge the way slides do, and arrive in the TED-talk voice unless you supplied a writing sample. The check: read the notes at speaking pace and mark every sentence you would never actually say — then rewrite those in the words you said while reading them, which are already tested for sayability. Fix the hedges ("this may suggest" becomes "this means"), and add the transitions if they're missing, because the bridge between slides is where presenters lose the thread, not within a slide. Finally, sync: if the edit pass changed a slide's claim, the note for that slide may still state the old one — notes drift from slides whenever slides get edited after notes were written, which is every deck. One search through the notes for the numbers that changed catches all of it.
The second-reader pass
After your pass, hand the deck to one colleague with three questions, phrased to be answerable in two minutes: "Where did you stop believing it?" "Which slide would you delete?" "What question do you still have?" These three catch what the author can't — you know the material too well, so every generic phrase fills itself in with your context as you read; a second reader hits the gaps as gaps. Skip the open-ended "any thoughts?" version of this ask entirely: it produces "looks great!" from a colleague who spent ninety seconds, and wastes the favor. Named questions produce named problems. And treat the answers as data, not commands — if two readers independently flag the same slide, it's broken; if one reader dislikes your strongest slide, that's taste, and the deck keeps it.
Editing someone else's AI deck
Half of all deck editing at work is reviewing other people's drafts, and AI raised the volume: more generated drafts, same review bandwidth. The workable version is a review pass, not a rewrite pass, and it starts with the title test — read their titles in order; if they don't form an argument, say that and stop, because line edits on a broken structure are wasted effort for both of you. Then two checks: the numbers ("which of these can you source?") and the generics ("replace 'a leading company' or cut the claim"). Fix as questions, not as silent edits — "is this 73% from our survey or from the AI?" — because silent corrections teach nothing, and the next draft repeats the problem. The reviewer's job is the deck being defensible in the room, and a presenter who knows which slide is weak, because you asked, defends better than one whose deck was quietly patched behind their back.
Editing for the room you can't see: async decks
Decks that travel without you — sent before the meeting, forwarded to absentees, attached to decisions — fail differently from spoken decks, and the edit pass needs one extra check. Read the deck as if you'd never heard the talk: every slide must be self-standing, because no notes are there to rescue it and no you to answer the confusion. That changes two passes. Titles carry more, because the reader skims titles and decides whether to read bodies at all. And slides may justly hold more words than a spoken deck — 80 words on a slide is fine for a handout if the words are load-bearing, because the walls-of-bullets rule exists to serve delivery, and silent reading has different physics. The test: send it to the one colleague who actually reads things in their inbox, with the question "what did you conclude, and what are you unsure about?" If their conclusion matches your ask, it ships; if they're unsure, their uncertainty is a slide you didn't write.
One slide, edited end to end
To see the pass at work, follow one slide through it. Draft: "Channel Performance Analysis" — a label title, eight bullets averaging eighteen words, a pie chart with six slices, and a conclusion hiding in bullet six. Pass zero: the title test fails, because the deck's argument needs this slide to say what's wrong with purchased lists — the title becomes "Purchased lists cost 3× per closed deal." Specificity: "significant cost differences" becomes the real $412-versus-$138 pair; "various channels underperform" names the two channels. Deletion: four bullets move to the notes, two merge into a table, the pie chart dies because the question was comparative. Verification: both costs trace to the CRM export; the "industry benchmark" the AI added gets deleted because it has no source. Read-aloud: the one surviving sentence is cut from twenty-two words to nine. Total time on one slide: four minutes. Multiply across ten slides, minus the slides that needed only two passes, and you have the 25-minute structure's origin story.
The 25-minute structure
Time-box the pass so it doesn't become an hour of arranging: five minutes on pass zero (titles, order, merges), eight on specificity and deletion, five on verification, four on read-aloud, three of buffer for the slide that always turns out worse than you remembered. Then stop. The deck won't be perfect; it will be defensible, which is what the room actually requires. Perfection is what the next deck is for.
What "done" looks like
Three checks, and the deck ships. The titles alone, read in order, tell the argument in under a minute. Every number has a source you could name out loud. And there's no slide you'd skip if the CEO walked in — which is also the fastest version of the test: flip through the deck and watch your own face for the slide you'd apologize for. That's the one the pass missed, and now you know exactly where to spend the last three minutes.
After a few decks you'll notice you fix the same three things every time — for most people it's hedging, generic examples, and one overlong slide. That's your personal checklist, the pass keeps getting faster, and the 25 minutes becomes the fixed cost of never being the person whose deck sounded like a machine wrote it. Because one did, and you caught it.