A few weeks ago I wrote about HealOS, the agentic operating system we built at Heal Capital. Today is not about that - because let’s be honest, many AI productivity gains are independent of which system you’re on. They’re in the little daily habits and the way you re-think processes.
Today I want to talk about the use cases and tools that personally made the biggest difference for me. I’ll add estimates of how much time it has saved me.
I’ll of course mention HealOS here and there. But you don’t need a custom-built setup at all for most of this. You’ll get 80% of the way there by plugging connectors into your favourite LLM account and building skills around them.
Let’s go, here are my top 5 points:
1. Talk, don’t type (30min/day)
The first one is simple: I’m an AI voice fanboy. I use Wispr Flow for dictation wherever I can: Sending messages on the phone while running from A to B, in the office to write emails or prompts, brainstorming things like this newsletter, and much more.
Once you do this enough, you’ll notice two things:
People think you’re slightly crazy. My Heal colleagues say it’s fine, but I know they’re still confused when I talk to my laptop. Also try doing it on a train… great social experiment
You need a specialized microphone setup… it’s just awkward shouting through the room, and Wispr makes more mistakes with a bad mic setup
Tbh I haven’t figured out which mic setup is best! Ideally you have something close to your mouth, that doesn’t look completely ridiculous. But a handheld mic doens’t work, it’s not convenient enough. And most other solutions make you look like a struggling insurance broker or like you just broke out of an institution:
Currently I’m simply using a lavalier mic clipped to my shirt - it works but it’s not perfect. I’ll most likely test one of those microphone rings soon!
They seem like the best mix of stealthy, hands-free and portable. OASIS (screenshot above) for example is a brand that has nailed the design, although they’re not shipping yet. These rings promise not only dictation but also smart navigation based on ring movement and finger taps. If that works, it would be a game changer for everyday agent interaction.
It would also be the first smart ring I’m wearing.
2. Stop clicking around your CRM (30min/day)
I rarely open our CRM anymore. If I want a visual overview of some companies, maybe. But over 90% of my use cases have shifted to a chat-based agent: logging companies, retrieving summaries, even sending emails to founders. I don’t want to open the CRM and click. I just wanna feel like talking to a smart colleague.
You might ask: But isn’t that making things more complicated? Can’t I just open my CRM and add a company with a few clicks?
The productivty driver for me here is chaining tasks. I’ll say something like: “Check out company XY, assess scope with our fund, set to initial screening, draft a first response, and remind me tomorrow to analyze their deck”. Or even better, the agent knows these steps are tied together already, just by dropping a company name. This process collapses five steps into one! The same applies for batch processing: I can drop a list of 10 companies into our agent instead of checking them one by one.
It’s less about replacing a click with a prompt, but rethinking how you interact with a CRM. In the beginning, a few clicks in your conventional CRM might feel easier. But you’ll leave a lot of potential on the table, so play around with it.
A crucial factor to make these frequent daily tasks work with an agent is latency. Waiting for an AI response for too long will feel annoying. The larger the model is, and the more connectors and tools are involved, the slower your response time. We’ve resorted to using smaller models (Haiku) and from a custom agent setup in HealOS. Faster responses are one of the advantages of using a custom system over using Claude/ChatGPT.
3. Find the hidden and painful admin cases (15min/day)
My favorite unlock with AI agents are stupid admin cases.
Simple example: When I’ve booked a flight, my agent pulls all flight confirmation from my inbox, blocks the flights in my calendar (with all links and info), and adds my commute to and from the airport based on the actual distances in both cities. The connection between email and calendar is just beautiful - be creative with it.
Another cool example I saw here is Blockit (Sequoia-backed), an email agent that automatically handles scheduling requests between multiple parties in an email chain. You just say something like “Thanks for sending your availability, Sharon. Blockit (in CC) will find a slot for us and send an invite” The pricing isn’t cheap though, so I might try and rebuild the use case with our internal agent.
4. Claude-skill-and-cowork-maxxing (0-2h/day)
An obvious one, but it needs mentioning! If you’re not using projects in Claude Cowork and Claude skills, what are you even doing?
Projects are incredibly useful if you want a shared working space across your team. I think they’re standard by now for most VCs to work on deals. Same for Claude skills - if you manage to build excellent ones and share them across team members, it’s a huge gain for your entire org.
Two skills that have saved most time for me:
Document generators for NDAs, one-pagers, investment memos and so on. It’s a no brainer. But I’ve discovered one useful tweak to increase output quality and reduce mistakes: My skills usually start with relevant questions to the user, instead of blindly drafting documents. The questions probe elements that need human judgement and hunt for blind spots in the data
A powerpoint skill trained on our fund’s style - colors, layouts and language included. I refined it over multiple decks with a feedback loop. Yep, unfortunately we still need to build tons of powerpoint slides for LPs and events!
That second skill might sound small but it’s very much needed… I can’t stand the AI-slop slides that Claude usually builds, they’re unbearable. Don’t use them for your pitch deck! Build a deck building skill instead that carries your own style and taste.
5. Building truly proactive agents (?/day)
This last one is not a small hack but more a mindset. Let’s be honest, most of us still use AI on a very reactive basis: Ask something in a chat, get one answer, repeat.
I’ve now committed to building agents that are proactive instead. One specific example: My early-deal-funnel agent helps me monitor all deals in the “initial screening” stage. Usually that means I had at least one interaction via email or call.
The agent works in multiple steps:
It catches inbound emails from founders that I missed, and suggests adding them to the CRM (=helps fill the funnel)
Creates a briefing for each company when needed, including what the last emails where about, timing and interaction history. It then asks me what to do next (=nudges me, gives me context)
Depending on my response, it then simultaneously drafts emails, updates the CRM status and adds notes to the CRM with the summary of my decision (=helps with execution and CRM discipline)
To be clear: Every email sent here still needs my approval, and the agent doesn’t decide - it suggests actions, fills gaps and keeps the process in one place. So no, you won’t get automatic AI slop messages from my agent.
The beauty is, these sorts of agents take the mental load off me. They help me close loops and stay focused on decisions that drive most value.
What has NOT worked?
Sometimes even better than knowing what works is knowing what doesn’t. I’ve abandoned the following AI use cases:
Daily or weekly briefings. I soon got tired of them - no significant efficiency or quality booster. The only thing that works surprisingly well for me is Granola’s call briefings. You get a smart briefing directly in the window where you’re taking notes. My conclusion: Briefings only work in the exact right context. Otherwise they add extra burden instead of reducing it
Fully autonomous email responses: We’re not there yet. I don’t even know whether I want it, but the truth is at the moment I can’t. Our agents don’t have enough context to nail each message 100%. Think about it, they’d need to know my exact communication style with that person and the full conversation history across all channels. I don’t want to compromise relationships by sending bad AI-generated messages.
Btw, Simon from Creandum has created a monstrous “AI brain” setup for this purpose, I highly recommend his article
Scoring or pre-sorting incoming deals based on an LLM and some system prompt… it just feels like you’re outsourcing your core job, and the result sounds like VC-slop. I’d rather eliminate all admin around the decision to zero and have more time to make these decisions myself atm
Sure, we’re likely going to move to AI-driven deal selection in the future as an industry. I’m not against that! Just saying you can’t rush a core part of your value chain. Especially not before you’ve addressed the admin burden.
Bottom line
All of these combined save me multiple hours per week. The biggest time saving on paper is creating long documents like memos or decks. But the greatest benefit might not be time savings. It’s the mental benefit I get when agents take my admin work and watch over my funnel. Feels like I can decide with full attention, without chasing emails or updating CRMs.
The biggest limit to my AI setup right now is context and memory. I need to give the agent better information on how the fund runs, plus my past interactions and decisions. Much of that is not documented anywhere (yet), or not in a format the agent can access.
We’re working on it. I’m sure I’m not the only one!
Speak soon,
Lucas
P.S. super curious how all of you handle this. any cool tools or hacks, especially for memory and context?