I’m using genAI tools for “messy middle” tasks

Part 2 to “What is useful genAI usage?” (6 months later)

A really important assumption that we need to keep at the forefront of our mind is that today generative AI models are NOT doing novel things or even thinking (see: basic addition). Instead, they are memorizing confidently. (Also they are getting better at memorizing things but they are not perfect yet!!; see: hallucination).

This assumption is crucial for many reasons, but it is particularly relevant when figuring out which “job” they are the right tools for, which I believe boils down to:

Making a huge amount of information understandable quickly.

Um, so not to drive this point home again / too hard, but what this means is that today these tools are not yet good at creative, out-of-the-box, net new thinking — even if they can eventually help spark that in you, a human.

There is a second thing that we should always have at the back of our mind, and that is model advancements. Last week, GPT 5.5 came out and it appears to be very good at complex, multi-hour autonomous tasks (with the right context). This costs lots of money to run, however, though maybe this matters less to you if you work at a Big Fancy Company.

However my suspicion is that because of the first insight and in spite of the second, the tasks you will get the biggest bang for your buck on with generative AI tools are not the very easy ones (searching for information) or the very complicated ones (hacking proprietary hardware), but somewhere in the messy middle AKA the not-basic-but-not-super-complex tasks. (A flavor of this task is design generation, but we will focus on that in a future piece.)

Below, I share more concrete detail about I am currently using this to help me in my day-to-day now that I’m six weeks into a new gig, but the main impact it has had is that it has helped me maintain my sanity¹. I use these tools for designing and building systems that speed things up by:

1. Letting me get quickly into the right details AKA focusing on the right things at the right time, while also

2. Making it far easier to make time to zoom back to do the fun deep thinking stuff product managers need to do

I have four key use cases currently:

1. Keeping my bar for the craft of product management high

2. Sussing out where to read carefully / probe when there is a lot of information to absorb and I’m still low context

3. Helping me deeply understand the user

4. Iteratively improving (as slowly as I need to) my stakeholder communications

Below, I go into detail of each of the above, namely why and how I do this, but honestly I would read it when you are in the middle of building automation for your tasks / this could be a good reference for future you.

---

1. Keeping my bar for the craft of product management high when I don’t have another person to do that with regularly, or if I’m worried people won’t challenge me

PM’ing is a lone wolf job. For example, most product management managers / leaders do not manage hundreds and thousands of people at big companies, they manage on the order of tens, and that is because there are far fewer PMs at a company compared to other functions in R&D. On top of this, they are not really in the trenches with other PMs, they are usually working closely with engineers, designers and data scientists.

This makes it really hard to get into constructive debates with your fellow product managers, so having a thought partner AI agent who is trained to do that is super helpful for:

1. Interrogating / steel manning arguments

2. Quickly connecting dots between different problems, or understanding how far along a solution is

3. When great samples are provided, crisply defining new problems that need to be solved, in the context of the market landscape (read: strategy)

What I’m using: Claude Cowork, integrated with Slack, Gmail, JIRA, Canva, and Snowflake via a data MCP. Context includes previously-saved gold standard PM craft literature and my sample documents and meeting summaries.

2. Pointing me in the direction of where to read carefully when there are too many sources of information

This one feels pretty obvious but the main thing I’ll highlight here is that you really should still go read the document, because genAI memorization isn’t perfect yet. I just use them as a way to help me sniff out where I should focus, rather than having to manually sift through so many resources to figure that out on my own.

What I’m doing: One agent I’ve built uses Zoom’s AI-generated meeting summaries which are actually really good and closer to the raw transcript, so I don’t do too much here. I just used Claude Code to build a daily job that pulls the summaries from yesterday’s meetings into my Downloads folder, and then formats and sends it to me as a slack message. This is not yet perfect because the (effectively CRON) job only runs when I’m online, since it needs to connect to the cloud where my Zoom summaries are, and you have to change your default Zoom settings to auto-transcribe which was surprisingly annoying to figure out how to do, but overall the system isn’t too bad.

3. Helping me deeply understand who my user is and what they care about

Canva users love this product more than any other product I’ve worked on and it is actually, truly a JOY — I love it when people are passionate / light up talking.

I am not, however, a: designer, marketer, PR person, influencer, yearbook editor, someone who needs a lot of creative tools all in the same place, or entrepreneur (primary target market). I fall into the bucket of: I make decks because I have to, and I work with words, and so my tools are paper, pen, workstation, docs (second target market). I do not understand what bothers them so deeply they would rather leave the software than keep trying if they can’t figure it out. For instance, last week, a fellow PM who was once a marketer told me about how frustrating it was to figure out which images have upcoming license agreement expiration dates, and wouldn’t it be wonderful if we could solve that problem for users? And there is no idea I would have even known that was an issue.

The challenge is that as a product manager, one of my key responsibilities is to deeply understand my customers. Like: how do they feel? Who are they - not their age or gender or the city they live in necessarily (though that could be useful, sure!) - but are they looking for joy when the world feels like it’s on fire, or are they just trying to get through their to-do list as fast as possible because they would rather be doing something else?

This unlocks making great decisions, quickly AKA the deeper I know my customers, the more I can imagine what they are going to need next even if they don’t know it themselves.

What I am doing: When you have an enthusiastic user base already, you have so much going for you! Online public forums like Reddit are usually active, users are down to give you feedback - often unprompted! - speak through their actions, like they will leave the product after wrestling with it so you can look at usage data, but they also speak directly AKA they also send in requests for help or proactively reach out. My AI agent collates all that information for me, using a lens of jobs-to-be-done, and provides me with a up-to-date dashboard and daily updates of top relevant complaints and feedback.

4. Patiently iterating with me on communications

This is probably the hardest part of my job - for me - AKA threading the needle effectively on communications when you are working with lots and lots of different *personalities*, usually virtually. It is further complicated when you are new to a company and new to a culture.

So what I have been doing is I jam with Claude on:

1. Interpreting engineering-speak

2. Prototyping future vision or other visuals rapidly for directional alignment with designers

3. Getting data independently to back or refute claims and to understand size and scope and complexity of problem, or as pre-work for discussions with data scientists

4. Brainstorming how to send a message given who the person is, and their role, and their level of tolerance for directness

This has a nice secondary effect which is it is in real-time training me to be a better communicator, almost like I have a private comms coach!!

What I’ve done: So here I use Claude Cowork or Chat depending, and Cursor. Since typing text is slow sometimes, I also brain dump using Wispr Flow - it’s really great at transcribing audio to text quickly and accurately (on mobile you can just use your microphone but it weirdly truncates things into several messages and isn’t as good at catching what you’re actually saying eg. BlackRock —> Blackhawk).

1. Honestly there is no way I would have been able to be on top of my new job at a new company while also moving to a new country otherwise (well I guess there is, and it would be that I would be broken / burnt out already).

Also pro tip: in month 1 nobody expects anything from you really except to listen so you can spend a bunch of time setting this stuff up and working through oh how come this integration with Gmail isn’t working yet, which is a lot harder to do starting month 2 when you are in the thick of meetings and five million slack channels.