LLMs are my exploited intern
How I went from from AI sceptic to having it summarize user research, build A/B tests, and more
When ChatGPT came out I read all of the LinkedIn posts about how amazing it was, tried it a few times, and dropped it. Answers were wordy and often wrong -— close to useful but never quite useful enough. It felt like LinkedIn influencers hyping their hot take over actual substance.
Listening to Marketing Against the Grain ~a year ago inspired me to give it another go. I realized that using LLMs day to day was going to be table stakes soon, if I didn’t do it I’d get left behind.
I started experimenting with incorporating ChatGPT and Claude into all of my daily workflows. At first, I tried to use it like a virtual assistant organizing my work but never got that to click.
What did click was when I switched to viewing it as an assistant to as an intern — taking on the tedious tasks I didn’t like.
Within a week I was a paid Claude subscriber. Now LLMs are my most used tool for work and personal projects. They’ve made work easier and even a bit more fun.
If you’re stuck at that useful but not quite useful enough stage with LLMs — this post is for you.
First, why learn to use LLM tools?
Don’t get left behind
Over the next few years (maybe already?) we’ll have “AI native” workers. They’ll have grown up with LLMs and use them instinctively. If you don’t develop these skills, you’ll be like an older office worker in the 80s shaking your fist at those new-fangled computers. Best of luck with that.
It’s faster
AI makes me significantly more efficient.
Depending on the task, it can range from a nice time saver (like getting a first draft of this post) to being exponentially quicker. And usually the time is saved on boring stuff I don’t like all that much anyway.
Expand your scope
We’re seeing designers become developers, developers copywriting, and marketers build their own tools. LLMs remove traditional skill barriers and let you venture into areas that were previously off-limits. If you’re not expanding your scope, you’re losing your edge.
Outsource the boring parts
AI handles the tedious parts—the data cleaning, the sifting through hours of interview transcripts, the boilerplate code. It frees me up to focus on more interesting work.
What LLMs are good at
Think of LLMs as an extremely eager intern with unlimited scope and energy. They can tackle an incredible variety of tasks and do them quite well, but need clear direction.
They are better at assisting with the doing parts of your work rather than the thinking parts.
Here’s my hierarchy of LLM usefulness, from most to least valuable:
1. Synthesising (most valuable)
This is where LLMs truly shine for me. They can analyse and sort thousands of records in seconds.
Example:
Give access to your anonymised user research interview transcripts. Share the background goal of the research, e.g. “discover user problems faced that prompted them to sign up to our software”. Have the LLM summarize answers with specifics prompts, “List problems user faced before signing up to our tool, rank by most mentioned problems”.
2. Collaborative coding
I’ve always coded, but never as a full-time developer. Previously, I’d make syntax mistakes or spend ages on Stack Overflow trying to configure solutions.
Now I can build functions and write complex SQL queries much more quickly. It requires back-and-forth iteration, but it’s still dramatically faster.
Example:
Writing SQL queries in BigQuery to combine Google Ads and GA4 data to understand attribution. Doing complex joins across multiple tables would previously have taken me hours. I was able to iterate with Claude (by sharing results and errors) and get it written in 30 minutes.
3. Creating written content
LLMs excel at first drafts that you can then edit and refine even if the writing feels a bit lifeless.
They help you get over the blank page problem but it’s worth remembering that good ideas are what make writing interesting. You’ll still have to generate those yourself.
Example:
Speaking the outline of an idea for a blog post (this one) to a transcript tool, uploading that transcript and getting a basic first version to edit.
4. Creating prototypes and web pages
LLMs are not limited to just outputting code, they can produce artifacts that you can download and use.
I used to rely on Balsamiq or Figma for wireframes. Now I often start with Claude. Sometimes the wireframe even evolves into a fully finished webpage without needing additional work.
5. Research (open synthesising)
Similar to 1 above, LLMs are quite good at research tasks where they synthesise the web or other large stores of data you don’t own.
I don’t have strong use cases for this in my daily work. But if you had to do a lot of market or other open research I can see them being very helpful.
6. Analysis & idea generation
LLMs aren’t all that strong here. It’s easy to think of LLMs as having infinite wisdom but at their core they are prediction machines. They use their training data to predict a good answer to your question.
I still like to feed Claude my project context and my own conclusions to see if it can wow me with an insight I’ve missed. I haven’t been truly wowed yet.
I am interested in connecting LLMs to tools like GA4 and Ahrefs for ongoing analysis.
I think where they can potentially help here is in ongoing monitoring of issues you don’t have time to dig into. Again the concept is that of an intern analyst.
How to get useful outputs
Provide context once
IMO, this is the most important factor for success.
Give your LLMs tool as much background as possible: your project details, market & company background, what you’re trying to achieve, relevant constraints, and success criteria. The more context you provide, the better the output.
I use Claude with Claude Projects, which automatically connects to Google Docs, GitHub, and other sources.
Having clear context in a project allows me to keep prompts shorter and specific to the output I want.
Prompt for outputs
Prompting is important—but think of it more as specifying your desired outputs rather than crafting perfect instructions.
Provide context and then use prompting to communicate what format, depth, and style you want as an output.
Conclusion
Claude won’t be making my coffee anytime soon but it has become embedded into my day to day work like a tireless intern.
For me it’s tactical rather than a strategic tool. I’d trust LLMs to synthesise existing user reviews but not to find truly creative solutions to what that research uncovers.
But with the right context and prompts it can dramatically accelerate your work. Conversion research that used to take days can now be completed in hours.
In my next post, I’ll show you exactly how I used Claude to redesign a SaaS homepage in a day - including the prompts that turned hundreds of customer reviews into winning copy.