Using Claude to Build a Pitch Deck That Sounds Like You

A workflow for founders who want AI leverage without losing credibility

Recently at a YC Discord meetup we were chatting about how almost every YC demo day deck is now Claude-generated. Even though it is very obvious, it still upped the quality of decks from previous years.

Back in 2022, when I was prepping for my own demo day, I remember our group partner telling us to do the literal bare minimum: include the necessary information and do not make the deck look designed or polished. His argument was that if you spend too much time on the deck, that signals you are not busy enough with the actual work. And if your startup is moving fast enough, your numbers should look like they are being updated every few days. Partly because of this mindset, we have companies like Airbnb with a seed deck that is bare bones.

That mindset made sense when polish was expensive. It no longer is. There is no dispute that even an AI design tool can do better than the average 22-year-old engineer CEO with no aesthetic aspirations. So, why not use it? Because the problem shifts to credibility. The suspicion becomes: if the deck is AI-generated, are you actually behind the story and the numbers, or are those also something the AI hallucinated?

This article is about leveraging the positive sides of AI to augment your capabilities without making the whole deck artificial to the point that your credibility gets questioned. I build pitch decks for clients using Claude as the primary drafting engine, and the output goes in front of real investors. What follows is the workflow: what to feed the model, how to constrain it, how to audit what comes back, and why the deck itself is only half the job.

The core mistake: one prompt, one deck

A pitch deck compresses months of a founder’s thinking into 12 slides. Asking a model to produce that compression from a two-paragraph summary is asking it to invent the missing 95 percent. It will, and everything it invents will sound like every other startup, because the model is filling gaps with the statistical average of every deck it has ever seen.

The statistical average of all pitch decks is exactly what triggers the credibility question. When an investor senses generic content, they stop trusting the specific content too, including your real numbers.

So the workflow inverts. You do the expansion first, then let Claude do the compression.

Step 1: Dump everything before you ask for anything

Before requesting a single slide, give Claude the raw material:

• A voice memo transcript of you explaining the company to a friend. Record yourself for ten minutes, transcribe it, paste it in. This is the single highest-value input because it captures how you actually talk.
• Real numbers. Revenue, users, growth rates, pipeline, burn. Exact figures with dates.
• Your origin story. Why you, why now, what you saw that others missed.
• Emails or memos you have written about the company. Anything in your natural register.
• Objections you have already heard from investors, customers, or skeptical friends.

Then tell Claude explicitly: “Do not generate content yet. Read this material, then list what is missing for a seed-stage deck and ask me questions.”

This step matters for two reasons. First, the model now compresses your material instead of inventing filler. Second, the questions it asks will expose real gaps in your story before an investor does.

Step 2: Lock the narrative before touching slides

A deck is a story with a slide count, so settle the story first. Ask Claude to draft the narrative arc as plain prose: one paragraph per future slide, no formatting, no bullets. Argue with the prose. It is much cheaper to fix a story in paragraph form than after it has been distributed across 12 designed slides.

A useful prompt at this stage:

“Write the pitch as a 500-word memo an associate would forward to a partner. If any sentence could appear in a competitor’s memo, flag it.”

That last constraint is doing real work. Generic claims survive editing because they sound fine in isolation. Forcing a uniqueness test surfaces them.

Step 3: Constrain the voice, explicitly and mechanically

“Make it sound less like AI” is a useless instruction. The model has no stable definition of what that means. Voice control works when the constraints are mechanical, testable, and specific. Mine usually look like this:

• Ban specific words and constructions. My standing list includes “revolutionize,” “seamless,” “leverage” as a verb, “landscape,” “unlock,” “empower,” and any sentence built as a contrast pivot (“it’s about X, and here’s the twist”). Add whatever your own allergy list contains.
• Provide a voice sample and name it as the target. “Match the register of the transcript above. If a slide headline could not plausibly come out of my mouth in a meeting, rewrite it.”
• Require claims to carry evidence. “Every quantitative claim must trace to a number I gave you. If you need a number I did not provide, insert TODO and ask.”
• Set a confidence rule. “State facts plainly. No hedging words, no hype adjectives. The data does the selling.”

Style constraints compound. Two or three produce mild improvement. Eight to ten produce text that reads like it was written by a particular person, because in effect it was: you specified the person.

Step 4: Run audit passes, one dimension at a time

Never accept the first draft, and never audit everything at once. A single “review this deck” prompt produces shallow feedback across all dimensions. Separate passes go deep on one dimension each:

1. Truthfulness pass. “Go slide by slide. Flag every claim that is exaggerated, unverifiable, or stated more confidently than my source material supports.” Models drift toward optimism when summarizing founder material. This pass catches it. An investor doing diligence will catch it otherwise, at a much worse moment.
2. Marketing language pass. “Flag every phrase that sounds like marketing copy rather than an operator describing their business. Suggest a plain replacement for each.” Run it twice. The second run catches what the first rewrite introduced.
3. Density pass. “For each slide: what is the one takeaway? Does anything on the slide fail to serve that takeaway? List cuts.” A seed deck slide should be readable in three seconds. If the takeaway needs a paragraph, the slide is doing the speaker’s job.

Each pass is a separate conversation turn with a single focus. The compounding effect over three or four iterations is what separates a usable deck from a plausible one.

The structure: steal YC’s baseline

You do not need a novel deck structure. YC’s guidance for seed-stage decks has been consistent for years: keep it to roughly 10 to 12 slides, make every slide legible from across the room, and put one idea on each slide. Kevin Hale’s version of the advice reduces to three words: legible, simple, obvious.

The canonical seed sequence:

1. Title. Company name and a seven-word description a partner can repeat.
2. Problem. Specific, felt, ideally with a number attached.
3. Solution. What you built, shown rather than described where possible.
4. Demo or product. Screenshots beat mockups. Real UI beats abstractions.
5. Traction. The growth chart. If you have one strong slide, it is this one.
6. Market. Bottom-up sizing an analyst could reconstruct.
7. Business model. Who pays, how much, and the margin math.
8. Competition. Why you win, framed as an insight rather than a feature grid.
9. Team. Why these people are unfairly suited to this problem.
10. Ask. Amount, runway, and the milestones it buys.

Add an appendix for everything else. When an investor asks a detailed question, pulling up an appendix slide signals preparation. Cramming that detail into the main flow signals insecurity.

Give this structure to Claude as the skeleton and let it propose which of your raw material maps to which slide. Disagree freely. The model’s mapping is a draft, and you know things about your investors that it does not.

A note on design, since Claude can also generate the actual file through code: resist decoration. White or dark background, one accent color, fonts large enough that the back row can read them, no stock icons, no gradients earning their place through vibes. Every element either carries information or gets cut. The same density pass from Step 4 applies to pixels.

Do’s and don’ts, condensed

Do:

• Front-load raw material: transcripts, numbers, memos, objections
• Lock the narrative as prose before generating slides
• Impose mechanical style constraints, including a banned-word list
• Run separate audit passes for truth, tone, and density
• Make Claude ask questions before it generates
• Keep the main deck at 10 to 12 slides and push detail to an appendix

Don’t:

• Prompt “make me a pitch deck” from a cold start
• Accept invented numbers, even plausible ones
• Audit everything in one pass
• Let the model set the confidence level of your claims
• Add design elements that carry no information
• Treat the first coherent draft as the final draft

The deck is a prop. You are the pitch.

Here is the part most deck advice skips: at seed stage, investors fund founders, and the deck is human enhancement material. It exists to make you clearer, sharper, and more credible in a live conversation. A perfect deck delivered badly loses to a decent deck delivered by someone who obviously owns every number on it.

Which means the same tools that built the deck should train the delivery.

Rehearse the narration with Claude. Paste the final deck content and ask it to play a specific investor: “You are a seed-stage partner at a fund that does healthtech. I will pitch you slide by slide. Interrupt with the questions a skeptical partner would actually ask.” Then answer out loud, in real time. The interruptions are the training signal. Investors do not wait politely for slide 10.

Drill the hard questions. Ask Claude to generate the 20 most uncomfortable questions about your business, ranked by how likely they are to kill the round. Write answers, then have it attack the answers. Two rounds of this and the live version of the question feels routine.

Use NotebookLM for rhythm and framing. Feed it your deck and supporting docs, then listen to the generated audio discussion of your own company. Hearing your pitch narrated back by someone else is strange and useful. You will notice which points land with energy, which sections drag, and where the story loses its thread. Steal the framings that sound better than yours.

Time yourself against the slide count. A 10-slide deck in a 20-minute partner meeting gives you roughly a minute per slide with half the time reserved for questions. Rehearse to that clock. If a slide consistently takes three minutes to narrate, the slide is overloaded, and you now have a precise signal about what to cut. Timing problems in rehearsal are density problems in the deck.

The loop closes here. Rehearsal feedback flows back into the deck, the deck gets tighter, the delivery gets faster, and after a few cycles the two converge on something that sounds like a founder who has thought hard about their company. Which is the entire point, because you have.

Summary

Claude will happily generate a generic deck from a generic prompt. Getting founder-grade output requires inverting the workflow: expand first with raw material, lock the story as prose, constrain the voice mechanically, audit one dimension at a time, and build on YC’s boring, proven 10-slide skeleton. Then spend as many cycles rehearsing the delivery as you spent polishing the slides, because the slides were never the product. You are.