How I use AI
Some tips from an addict
Every so often I’ll be talking to someone about AI. And they’ll say something like “oh, I tried it a year ago and it didn’t work very well so I’ve not used it since”. Or perhaps “I don’t really know how to use it”. Or sometimes “it refused to do what I wanted so I gave up”.
I get it. Learning how to do something new is hard. It requires energy. It’ll probably take more time than a familiar way. But AI can be magical. It saves me many hours each week. It helps me do things I couldn’t before. Here are my top tips and tricks.
You need to be a little careful - some organizations have rules about sharing internal confidential information with non-approved LLMs. But, in reality, there is much that all of us do which is non-confidential and where an LLM can help us.
#1 - always try to involve it
Whatever you are doing try to involve an LLM. If you are writing a doc, give it the skeleton and get it to suggest structural improvements. If you’re fixing things point the camera at it and ask for help. Ask the model to ask you questions to find out areas where it might be able to help. Tell it to be imaginative.
You can produce MP3 files from text (not sure about the robotic voice though…)
Or convert hand drawn charts into neat versions - this scribble…
…becomes a neat chart.
Or get dietary advice:
Although the answer changes if I tell it I’m a robot…
Get it to tell you where you are:
(It’s actually the Pentlands in southern Scotland).
Or convert hand-written notes to text. Or take a picture of a recipe and scale quantities (or convert from ounces to something more modern :)). Or convert speech to text.
#2 - if you’re stuck… …ask it
If you’re wondering how to use it, then ask the model how it can help. If it responds with a wall of text, then tell it to be concise e.g. ‘two concise answers only’.
And if you get writers block and don’t know how to start or continue a paragraph - ask the model to complete it for you. I rarely use the output verbatim but I find it’s great at unblocking me.
#3 - if it refuses then reassure it (or be impressed)
Here are the lyrics for a nonsense song GPT4 wrote. Claude initially refused to extend the lyrics. But when I told it I’d written them it went ahead (I like its justification - it’s OK ‘since these are your original nonsense lyrics’ :)).
So if the model initially refuses then ask it again. Sometimes asking nicely works. Other times a simple ‘try again’ can get the job done.
Sometimes I’m impressed. I asked Claude for a script for cold calling. It (correctly) refused. Claude is pretty lawful.
ChatGPT is less lawful. It happily gave me a cold calling script. Which I promptly binned :).
#4 - don’t overthink prompting; iterate instead
Prompt engineering was a thing in the early days of AI. More recently it has faded as people have realized that iterating is better than trying to craft a magical prompt that gets the ‘correct’ answer in one shot.
Most of my exchanges with the AI involve back and forth as I tweak my request and guide the model. For example…
And eventually we got there!
#5 - get the model to ask you
Models are trained to give their best answer to the question you asked. But often our prompts are unclear - the model needs more information to do a good job. In those cases the model should ask questions to get clarification, but by default it won’t. Telling the model to ‘ask any relevant clarifying questions’ can be useful.
For example asking GPT4 to review a doc triggers an immediate review…
But telling the model to ask clarifying questions can steer the review and make it more useful.
#6 - get the model to review its work
Models generate output sequentially and most don’t have the ability to go back and ‘edit’ the output. The only one that can is o1 - simplistically the reasoning process can be viewed as the model reviewing the output it has generated so far. Asking non-o1 models to review the output is akin to a basic version of o1 - and can result in noticeably better output.
#7 - use the best tools
By default I primarily use Claude for day-to-day tasks, with GPT4o for a second opinion. o1 is my current go-to for coding. I use GPT4o for images. Copilot sticker creator is also pretty neat. And Google has a neat “make a plushie from this picture” feature.
Of the others:
Six months ago Copilot was pinned to my desktop. But these days it’s feeling a bit dated despite the new UI. It disappeared from my taskbar in a recent Windows update - and it’s probably telling that I’ve not gone to find where it went.
The open-source models are interesting as I can run them locally. But they aren’t nearly as good as the frontier models. If Claude cost, say $500 a month, then maybe I’d explore them more. But Claude is a bargain at £18 per month - and I suspect the free plan is probably fine for most people.
I paid for Midjourney. But the UI - via Discord - was so bizarre that I quickly went elsewhere.
#8 - cross check between models
If I’m writing code, I’ll ask Claude, GPT4 and o1 to review code. With docs I’ll ping them back and forth between the models. Different models seem to come up with different ideas. Claude is very lawful and will spot ways I might be getting myself into trouble. GPT4 is a bit more gung-ho.
#9 - get multiple suggestions
If you’re asking the model to generate something then why not ask for multiple versions?
#10 - remember the context
All the messages in a thread are processed by the model each time you re-prompt. So things mentioned earlier in the thread can impact the later thread. For example:
Why is ChatGPT using Scots? Because earlier in the chat I told it this:
(As an aside you can see cross-context GPT memory at work - that’s why ChatGPT is asking about holiday planning).
#11 - it’s less distracting than Google
I’m surprised this one hasn’t got more attention. The UIs for Claude and ChatGPT are basic. And dull. They answer the question you ask. They don’t (yet) contain banner ads. Or pop-up videos. The answers don’t contain product placement or thinly veiled attempts to sell you things.
Nor are they deploying psychological tricks to keep your attention.
I’m pretty sure there’s a productivity boost just from being less distracted. Long may this continue…
Try, try, and try again
I’d be lying if I said it wasn’t frustrating at times. This morning GPT4o drove me mad as I tried to talk to it. It was forever interrupting to say “That’s a great idea” or “You’re on the right track”. But each time I learn something. And every so often I find a way to use it more effectively.
Think of AI as a cross between Einstein and a puppy :). Enthusiastic, positive and incredibly knowledgeable. But it needs guidance - you have to adjust to it since it can’t adjust to you. Keep experimenting, keep exploring, and you might be surprised at what becomes possible.
After all, if at first you don't succeed...