Jul 09, 2023

Writing damn good decks with LLMs

Hey there 👋 I'm building CodeTrail, which helps engineering teams document and share knowledge close to the codebase with no friction. If you're still using Notion, Confluence or Google Docs to document your engineering work, give it a try and let me know what you think!

I’ve always loathed working on tasks that don’t create actual value for anyone, such as preparing read decks for potential investors. You know that nobody really takes their time to read what you spent days of your life working on, but traditional institutions and government programs with a low propensity to risk-taking want to offer resources to startups, so you better come up with a business plan and read deck. The content is usually strongly based on requirements such as market sizing (TAM, SAM, SOM with CAGR), as well as other carefully-researched (read: 50% market reports and 50% made up) metrics so far from reality, they could as well be omitted.

And yet, every couple of months I find myself preparing another of those decks, contemplating the meaningful tasks like customer validation I could have used the time for. Luckily, with the advent of LLMs things have changed.

Recently, a repost of a news article on LinkedIn of all places caught my attention, revealing that GPT-generated pitch decks consistently outperformed their human counterparts in terms of quality, thoroughness, and clarity with 80% of respondents finding the decks compelling while only 39% of respondents felt the same way about human-created decks.

This raises an interesting question: Why bother creating decks manually if the probability of success is significantly lower at almost half the rate of an AI-generated deck?

In the following, we’ll walk through a simple workflow I’ve used for recent decks, which have been remarkably better than what I came up with previously.

Preparing the most important points

While most of the subsequent tasks will be more or less taken on by GPT4, we still need to make sure the model understands what we’re intending to end up with. While we could probably generate everything including the business model, there’s a good chance you’ll want to do this yourself. Jot down a couple of bullet points for the target customer, problem, solution, and whatever else you’re required to submit, and don’t spend too much time.

From bullet points to slides

Once we have our core bullet points, we can combine those with the list of requirements, and ask GPT4 to create slides for the deliverable at hand. For a read deck, I asked to generate read deck slides for the specific submission I was working on and included the guidelines and my content. Optionally, you can generate image suggestions to go with the slides, or whatever else you come up with.

After this step, you should have your slide content, as text. Before you move over to your slide software of choice, you can improve the slides to match all requirements.

Improving slides with a feedback loop

Since transformer-based LLMs tend to make stuff up and follow the instructions more or less, it can be useful to run a feedback loop, instructing GPT4 to criticize and improve the content you just generated based on the requirements. This usually spots all the missed areas and recommends changes to fill the gaps. Run this a couple of times and the most important requirements will be met.

Elevator pitches, no more

If you ever need to get creative for a deck, feel free to ask GPT4 to come up with elevator pitches, brand names, and other creative content that you’ll likely discard immediately after submitting the deck.

What does this mean for investors?

Similar to other recent developments, I predict that most founders will use GPT4 and related tools as copilots to create better decks and kickstart the creative process, while a small minority will take a shortcut and generate decks completely with AI. This will lead to an increase in seemingly high-quality applications at popular investors and accelerators, which have to find the needle in the haystack. A scenario where potential founders generate tens to hundreds of different AI-generated decks for ideas generated with GPT sounds like an arbitrage strategy until investors manage to do proper due diligence (think about the upside of getting term sheets for $20/mo of ChatGPT Plus). Investors will probably flock to the same tools, so we’ll end up with what my good friend Tim fittingly described as a future of AIs reading and filtering content generated by AIs.

Eventually, this development may lead to a desirable effect: Since decks no longer give you any guarantee to decide whether an investment is reasonable (they never did), decks might be replaced by other forms of due diligence. Personally, I hope that we’ll move more in the direction of face-to-face interviews with founders because that’s the only important component in an early stage deal that matters. Ideas come and go, and problem statements change with ICPs, but the founding team has to be capable of executing their way to product market fit and beyond. Artificial requirements like business plans have been a nuisance for far too long, so I welcome the advent of generative AI for meaningless tasks.

Thanks for reading this post 🙌 I'm building CodeTrail, which helps engineering teams document and share knowledge close to the codebase with no friction. If you're still using Notion, Confluence or Google Docs to document your engineering work, give it a try and let me know what you think!

Bruno Scheufler

At the intersection of software engineering
and management.

On other platforms