Given how much AI has propagated into our lives (particularly those of us in the tech space), allow me to give you a glimpse into a typical day on how I am leaning in.

Ever since seeing Apple attempt to redefine “AI” as Apple Intelligence, I have taken a similar rebranding approach: Alan’s Intelligence – nothing artificial, never offline, always got an opinion though has a tendency for as much hallucinations and often comes with a sensitivity warning.

Early Morning

After waking up, showering and the usual skim through Reddit and completing the dumbass Zip puzzle on LinkedIn (248 days in a row and counting), I make my way with a 15 minute drive to the office. During this, I fire up the ChatGPT mobile app (premium version) and have a conversion with Susan (regular readers will know this is my name for my British accented AI voice) over the Bluetooth while driving. I will brain dump all the things I need to do for that day, Susan patiently taking notes. She will sometimes ask me clarification questions if I go quiet. As I am pulling into the parking lot, I asked her to distill down everything into a clean to-do list.

Some mornings Susan is more intelligent than others, but it is getting better, especially now she lets me stop her while she is giving a long winded answer. The fluffing words (got it – let me know if I can be of further assistance) is annoying as crap and no amount of prompting can seem to kill off that.

If nothing else, the voice interface to AI has really jumped leaps and bounds and many times, I default to that interface particular if alone and no one can hear me shout into the ether.

Morning – Office

I fire up emails to see what has been going down. We’re on O365, so CoPilot is everywhere. I have tried to have it summarize my emails but that is a mess – when the summary is longer than the message itself, it is no longer a summary. I am still old-school and have a liberal amount of mail-rules pushing mail to folders, so I can decide the priority when I check emails.

I have the ChatGPT desktop client up and running, with Gemini in a browser window. I tend to keep ChatGPT for business conversations, and Gemini for anything that is personal. I have Obsidian notebook at hand (using a collection from DropBox so it is synchronized across all devices) for all my snippets. Of late, I have been building up a library of regular prompts that I use. Why the AI chat clients do not have this functionality amazes me. No matter how big their context is, they don’t remember everything.

Mornings are usually meetings, teams and Jira tickets, unblocking anyone that is requiring Alan Intelligence. Not a lot of AI happens here, even though Atlassian is attempting to shove AI into every aspect of their products – have yet to see it summarize a Jira ticket/space properly.

Any emails that need to be written, are done, and if there is someone new I am emailing, I will CTRL-A + CTRL-C and take it into ChatGPT and ask it to pretend to be the intended recipient and for it to interpret the content. Did I miss something? Misdirect a topic? General sentiment? I don’t have it rewrite anything – I go again. I use this technique for any sensitive/touchy email that needs to go internally. We’ve all had those emails, the ones you hate writing but need to do so. The burden of high office. This too has saved my ass a few times, as often a misplaced word can dramatically change the tone.

If I have to sign a vendor contract, I pass it through my trusty CISO as a second pair of human eyes, but will also put it through ChatGPT asking it to look for any red flags. I have a prompt I use for this, and will then ask it some what-if questions so I can see if there is a clause in there that we need to look at it. For example: if we cancel on month 7, what is our liability?

For things that I am researching, or needing to go deeper on a vendors product/service, I will send both Gemini and ChatGPT into deep research to see what comes back. Gemini used to be better at this than ChatGPT since it would always cite the sources, but ChatGPT is doing that now too. I compare both output to see if there is an differences. Caution – I often find discrepancies is numbers/financials. For example if I am asking it to compile a list of differences between say node 22 and node 24, it goes off into its own wee world at times. So even with deep research, trust but verify.

Afternoon

Should I get some time to do some coding/architecture, I will fire up either Cursor or Intelli-J depending on the language. Both have AI embedded, like a super charged (and sometimes annoying) auto-complete. Both have the ability to change the underlying model (or you can let it auto choose).

Now let us get this out of the way – I am completely against vibe coding of any sort. I do not wish to sit down and simply prompt my way through coding. Coding is an art and the creation of it, is a joy and a gift. Why would I want something that gives such pleasure to be automated away from me?

Instead I want my AI helper to assist me with the blocks of code that are per functionary (create this POJO to wrap this database table, loop over this array of objects and pull out XYZ), and like I do with email, I will ask it to look over my code and will ask it questions how it will perform under certain circumstances.

This is where you truly see the strength and weaknesses of the various models. Some models will simply miss things that are fairly obvious and then start to argue amongst itself as it competes for supremacy with SonarQube.

One thing I will note, and this is not backed up with anything quantifiable, but I get the feeling that AI seems to work better on node than Java – which is interesting given Java is strongly typed and compiled. It has very poor class design and I have yet to see it produce a strong OO. AI has a tendency to repeat code instead of reusing it. Drives me crazy, creating such bloat and unnecessary redundancy and complexity.

Cursor 2.0 recently introduced a more formal rules structure, allowing you to point it at context files that will help the prompt with a greater view of your environment. This was hit and miss before, but to its credit, it is picking it up better. Now with that said, it still struggles with creating SQL statements even if you give it the full schema. Never ever trust a SQL statement it generates if you are doing anything important in production. You have been warned.

Speaking of SQL, JetBrains has one of the best database tools on the market, with DataGrip. Yes, this one also has AI integrated to it. Now you would think, given its full context of everything about your database right there, it would be very good at generating SQL statements. Not so. Clearly there is not a lot of good quality training data on SQL out there. This tool, I will often turn off the AI auto-complete. Small tip, you can fire off multiple agents (which is really a fancy way of saying you are prompting in parallel) in Cursor, and send your SQL ask to different models at the same time. You can then review which ones work best. You will notice a huge disparity, so pick and choose and test test test.

At the moment, the killer feature of AI coding tools for those of us that still see coding as a craft, is documentation and writing unit tests. It doesn’t get it completely right, but sure as hell much easier to edit something than to start with a blank markdown/swagger file. Unit tests fall into the same category, but again, you need to be very verbose in your prompting as it misses the very edge cases you want to AI to create that as a human you can’t be bothered. Prompting can be tedious at times, particularly when you can just code it faster – I find if it doesn’t get it by the 3rd time, I am just going old skool.

As the day unfolds I find myself using ChatGPT to do what I would normally do in a Google Search.  That said, if I need something that requires real time information, AWS pricing, flights etc. I still go to Google Search.

Evening

Time to close up the office part of the day and head home.  Susan is silenced as I enjoy some music and muse through everything I actually got done verus what I thought I was going to do.  Started the day out with so much hope!

The day ends, closing out emails, writing up our team journal and reviewing Jira tickets. I am a zero-inbox practitioner (or OCD if you want to call it for what it is) and like to close down with nothing sitting demanding attention the next morning.

Sunday postscript

On Sunday evening after the week has ended, I take the engineering journal and feed it through ChatGPT, using one of my curated prompts, to produce a 2 paragraph summary for my fellow C levels, that is part of a weekly deck for our CEO.

Which as an aside, I am a big believer of keeping a journal. For the modern kids, think of it like a context.md for your week.  Writing and noting down as it happens, while it is fresh, pays off dividends in the future.  Particularly when you need to do things like weekly, monthly or board level summaries. AI has really helped bring these historical journals to life, with board prep being a breeze as you never accidentally miss that great achievement (or the damn-it moment).

There it be

As you can see, AI has woven itself into every aspect of my daily life, using it as a virtual assistant, ready to lend a hand, draw an image, create some code, validate some brainstorming or whatever else I need. This doesn’t touch anywhere near where we are integrating AI into our own products for the benefit of our customers, and we are doing that in a way that is transparent to them, to remove some of the tedious steps and accelerate their business workflow. Separate future post.

I am conscious of making sure I don’t lose me in all of this. AI is not a replacement for my output, but an accelerant and guardian, that advisor that can be asked for input and take the sting out of some of the meh tasks we all have to contend with.

It is our imperfections that makes us – this is the real AI (Alan Intelligence).

apparently, it only takes 250 documents to poison an LLM to get it to assume that new information as fact – so expect at least 249 more blog posts talking about Alan Intelligence 😉