How I'm Using AI

Adventures with ChatGPT, Claude Code, Gemini, and more

Every time I go on Twitter / X, I am bombarded by massive waves of FOMO. Everybody seems to be using more AI better than I am to automate everything about their lives. People claim to have hundreds of agents running continuously, automating everything from product development to social media management to customer or founder outreach to email responses, code generation, blog writing, and even financial modeling. This feeling has become particularly acute over the past month, as recent models releases from Anthropic have meaningfully improved the power and capabilities of AI agents to do ever more work for humans.

So, to hold myself more accountable to experimenting more with AI tooling, I decided to document how I’m actually using AI today and how I’d like to use it more in the future. I want to “build / learn in public” and gather feedback on how to better use AI to improve my life (so if you have ideas on more I should be using AI for, please reach out and let me know).

Today, I’m using AI in a few main functions:

1. Market research

2. Company diligence

3. Podcast editing and promotion

4. Writing support

5. Financial support

6. Other backoffice

7. Fun :)

The main AI tools I use today (more or less in order of how much I use them):

1. OpenAI’s ChatGPT

2. Anthropic Claude (normal Chat and Code)

3. Dia browser (although, honestly I’m considering switching back to Chrome + Claude Chrome extension)

4. Obviant

5. OpusClip (to generate podcast “shorts”)

6. Riverside FM (for podcast editing)

7. Otter AI (for podcast transcription)

8. Gemini (built into the Google Suite)

9. Perplexity

10. Tegus AI features

11. Canva AI features (I only recently discovered this, so TBD how much I will continue using this)

Market Research

Whenever I’m looking to learn about a new market or technology area (whether for company diligence, a blog post, or a podcast), I always generate a “Deep Research” report from ChatGPT. These reports are essentially customized Wikipedia articles and tend to surface high quality “explainer” articles on certain technology / market areas. I particularly like Deep Research because I can tailor my research request to my own niche interests (ex: looking at how a tech area specifically relates to startups or national security, which your average Wikipedia page or explainer article may not dig into). Overall, similar to Wikipedia, I view Deep Research as a good starting point to learn about a space at a high level, but definitely not the end destination.

I use ChatGPT and Perplexity for smaller research questions (ex: specific questions about how much a certain company has raised, to list relevant companies in a technology area, to find a specific contract won by a company, etc). Recently, I’ve found myself using Perplexity less and less as ChatGPT’s search functionality has improved.

While researching a subject, I use AI to help parse long documents and identify key sections of interest. To do this, I typically use Dia’s AI-native browser sidebar chat, which allows me to work directly within the browser without downloading large documents or re-uploading them to a separate chatbot. For example, I heavily used AI to identify key parts of the NDAA, Appropriations Bill, and Joint Explanatory Statement that were relevant to startups and innovative technology developers in order to prepare for our podcasts on the subject.

I use Obviant’s built-in AI features to quickly get up to speed on Department of War (DoW) market opportunities (ex: search the J-books, search for major programs / contracts, etc).

Whenever I need to analyze large data sets, I use ChatGPT or Claude to generate Python scripts to do that data analysis for me (ex: I used ChatGPT generated Python code to analyze the data I discuss in my blog post analyzing Navy contracts).

Company Diligence

A major part of my job is conducting “diligence” on potential investments – that is, deciding if a company is a good investment or not. Most of my diligence process remains pretty manual and human-driven (ex: doing reference calls on founders, speaking with customers, speaking with competitors, spending time with founders to understand how they operate, thinking deeply about a market and technology approach), however, AI is a good thought partner and research assistant during the diligence process.

I use ChatGPT and Perplexity to identify competitors and relevant valuation and exit comparables (ex: identify all companies in a similar market area that have been acquired, gone public, or raised money). I also use both ChatGPT and Perplexity to help with market sizing and broader market and technology research (as described in the section above) as it relates to company diligence.

I use Tegus’s internal AI chatbot to search their transcript library to identify relevant interviews relating to market size, customer spend, and competitors. For example, Tegus transcripts can show how much customers budget for a product, what they spend on competitors, what they actually pay for a particular company’s offering, where customers see major gaps in the market, and what they like and don’t like about their current tooling. Honestly, while I love Tegus, its AI product is only okay – I hope they improve it in the near future.

I still write most of our diligence memos manually, as writing helps clarify my thinking on a certain company / market. However, I will use AI to generate founder bios based on their LinkedIn pages (again, I typically use Dia’s built in chat agent in the sidebar to do this). Sometimes I will use ChatGPT to describe a company’s product based on their pitch deck.

Occasionally we need to share diligence memos outside of the firm. I use ChatGPT to “polish” prose in diligence memos for clarity (typically our internal diligence memos are pretty messy, with lots of off-the-cuff notes and thoughts, often in the form of bullet points).

I use Obviant’s AI search features to look up companies’ DoW contracts, as well as those of their competitors, and to search the budget and J-books to understand the scale of the DoW market opportunity for a particular company.

Podcast

I record all my podcasts in Riverside FM which has a number of AI editing features including the ability to:

• Automatically generate podcast transcripts

• Suggest parts of the podcast to cut

• Identify and cut filler words and long pauses

• Suggest ideal “shorts” to post to help promote the full podcast episode

• Automatically tune audio to improve sound quality

• Includes a chat interface that can be used edit the podcast using natural language

In general, I don’t take huge advantage of Riverside’s AI features beyond its transcript (we have an amazing human podcast editor who does most of the editing for us), but I have used the features occasionally when I needed to get an episode out quickly and didn’t have time for someone to manually edit an episode (ex: I did this for our emergency pod on the War Department’s new AI strategy).

I use OpusClips to automatically generate podcast shorts (I think OpusClips does a better job than Riverside in clip generation). OpusClips identifies the 30 most clippable sections and automatically generates live captions for those shorts. It will also automatically generate titles and social media post suggestions (although I don’t use those very frequently). Normally, I use ChatGPT to generate Twitter / X, YouTube, and LinkedIn captions that accompany the shorts, which I publish as a “drip campaign” to help promote my podcast episodes.

I’ve recently started experimenting with using Canva’s built in AI features to automatically generate other marketing collateral for the podcast, such as LinkedIn “carousels.” I’ve been pretty impressed with this functionality – I was able to upload my podcast transcript to Claude which then automatically generated 5 slides for a carousel which it then sent to Canva (via Canva’s Claude MCP connection) to generate. While the generated carousel isn’t amazing compared to what human designers can do, it’s pretty good (and certainly better than anything I can come up with myself).

I also use AI to help structure interview questions. Typically, I manually brainstorm questions, then I ask ChatGPT to make the questions sound more conversational and to order them into a good conversational “flow” – ChatGPT typically organizes the questions into thematic buckets and suggests an ideal order to ask those questions. Similarly, when I moderate in person panels, I’ll use ChatGPT to research the panelists and generate thematically relevant questions for each panelist based on their backgrounds. I also use ChatGPT to generate conversational sounding introduction scripts for the podcast based on notes or a first draft I provide it.

Writing Support

I do the vast majority of my writing manually, as the writing process helps me clarify and structure my own thoughts. However, I will use ChatGPT to 1) help me re-write clunky sentences that I just can’t figure out on my own, 2) write simple definitions for jargon (ex: most of time if you see a footnote providing a definition for a word, it is ChatGPT generated with light editing from me), and 3) review drafts to identify typos, grammatical areas, or ways to strengthen the piece (ex: identify areas where I use passive voice, suggest areas that are repetitive and could be consolidated, etc).

I will use ChatGPT to help me come up with good, “clickable” titles for both blog posts and podcasts.

Financial Support

Despite technically working in the finance industry, I do not actually do a whole lot of financial modeling on a day to day basis, and I have no formal finance training. Occasionally, I come across complex financial scenarios I need to model. ChatGPT (particularly in thinking mode) is amazingly helpful when it comes to financial modeling. In general, I don’t have it fully create spreadsheets for me, rather, I ask for help thinking through the math needed to model particular scenarios (ex: I had to model how a certain kind of warrant would affect a company’s cap table, which I had never done before, and ChatGPT helped me figure out the formula needed to do so). ChatGPT also helps me figure out how to run certain commands in Excel (ex: how to sort lists in a certain way, do particular calculations, etc).

Other Backoffice

• I make liberal use of Gmail’s AI autocomplete functionality.

• I use ChatGPT to generate small images for event links like Luma and Partiful.

• I recently hooked up Claude with an MCP server that connects to Affinity (via Zapier), which lets me update Affinity through Claude’s interface. The connector is pretty unreliable at the moment, and I’m working on building an internal app that just uses Affinity’s API rather than messing around with MCP

• I just started using Google Slides’ “Beautify this Slide” feature which I’ve been pretty impressed by. I am not a good slide designer, but the “Beautify” feature uses Google’s Nano Banana model to generate an interesting slide design based on text bullet points I put on a slide. It still has that “AI-generated” look, but it’s better than my basic bullet points.

• I recently vibe coded an app that automatically uploads a pitch deck to Google Drive and also adds the company to our CRM Affinity. Previously this was a pretty clunky workflow where I had to upload the deck to Drive, copy the link to the file, add the company to Affinity (which itself has a pretty clunky interface), and then paste the link to the file into the company’s Affinity profile. I can use this both as a web app or as a Claude “skill.” Next I want to add the ability to automatically upload Notion notes to Affinity profiles (again, today I need to manually copy and paste the notes in, but I’m sure there’s a better way…)

Fun Stuff

• I use ChatGPT to generate recipes (one highlight: over Thanksgiving it generated an extra moist pumpkin coffee cake recipe which was incredible) and ask for other cooking advice

• I’ve been using ChatGPT for workout planning and advice. For instance, I’ve been trying to do more lifting this year (without injuring myself) and I’ve been using workout plans provided by ChatGPT. I also ask questions mid-workout about how to improve my workouts. I’m planning to do another half marathon in the spring, and I will definitely use it to train for that

• I am horrible at keeping house plants alive. I recently moved and bought a bunch of new house plants, and I’ve been using ChatGPT to help me take care of them

Things AI Still Isn’t Amazing At (in my experience)

I’m very excited about the future potential for browser and computer use agents (that is, agents that can actually click around and use a browser and computer the same way humans do). The browser use agents I’ve experimented with (mainly Claude) are still extremely slow and clunky and seem to struggle to do basic tasks. They still rely on “screenshots” to understand what is happening on a website, and they lack basic functionality like uploading a file to a website (file pickers are at the OS layer, not the browser layer, so Claude’s browser agent can’t access files). This is frustrating, as most of the work I do is in web applications.

I’m also struggling to get agents to work through more complex mutli-step workflows. For example, one workflow I would like to automate: upload a full podcast recording, have an AI choose the 5 best shorts, have an AI generate those 5 shorts (with live captions), have an AI draft captions for each short tailored to LinkedIn, X, and YouTube, then have an AI schedule those shorts across those social media platforms to post one a day for a week. This is something that is easy for a human social media marketing manager to do. However, when I asked Claude to do this workflow, it got stuck in the middle. Impressively, it was able to identify the 5 best shorts and generate those 5 shorts, however, it was not able to generate live captions for those shorts or schedule them on social media. I know many AI expert users recommend breaking up complex tasks into multiple smaller steps.

I’ve also found MCPs to be pretty unreliable and have started just using Claude Code to build small internal apps that leverage 3rd party APIs rather than MCPs.

Looking to the future

I know that I’m only scratching the surface of what can be done with AI to make me better at my core job: finding and supporting amazing founders building critical technology for national security customers. In particular, going forward, I want to spend more time using coding agents to develop some custom apps – beyond a few toy projects, I haven’t spent enough time experimenting with the art of the possible here, particularly in light of advances in coding agents like Claude’s Opus 4.6 and OpenAI’s Codex. I’d like to build some apps that let me do things like: automatically parse a pitch deck and upload all information to our CRM with all necessary fields filled out, automatically analyze large numbers of Tegus transcripts to pull out key findings on relevant companies and technology areas, and so on.

I still believe that the core work of early stage venture is fundamentally human driven, and unlikely to be taken over by AI any time soon (although only time will tell). Ultimately, early stage investing is all about the people you’re investing in – meeting them, building relationships with them, understanding what drives them, and then supporting them. But I do believe AI can give me more time to focus on what matters (actually spending time with people) and make me better equipped to understand the companies and technologies these founders are building.

Please let me know what I’m missing out on! If you’ve found interesting ways to use AI to improve workflows, please tell me, as I’m actively trying to learn and improve my current use of AI. And as always, please reach out if you or anyone you know is building at the intersection of technology and national security.

Other people who have written about how they’re using agents that I found interesting

• Get Good at Agents by Nathan Lambert

• Use Agents or Be Left Behind? A Personal Guide to Automating Your Own Work by Tim Dettmers

• Could Claude Code Work for ChinaTalk? by Jordan Schneider

• Claude Code is the Inflection Point by Semianalysis

Note: The opinions and views expressed in this article are solely my own and do not reflect the views, policies, or position of my employer or any other organization or individual with which I am affiliated.