# Casework, Code and Lash Extensions: How I've embraced LLMs, and my limits > Source: > Published: 2026-09-17 15:11:04+00:00 I know this is a controversial subject, but rather than not talking about it publicly for fear of cancellation, I do want to add my voice to the current talk about AI tools, and share how I’m using them: professionally, politically, and personally. Earlier this year I took an extended period of medical leave to recover from surgery, and upon returning to work after almost 3 months off, I felt like the tech industry had hugely shifted in that 3 months, with AI tool adoption now huge. In that time I also moved from being a senior software engineer on a team to instead being the tech lead for the entire public product portfolio at my employer. When I came back, I wouldn’t be writing code. I had a bit of an existential crisis about this: were my skills already obsolete, would I be left behind? So, whilst recovering, I took the time to work on a personal project (the website for my wedding later this year) using Gemini Code Assist (as a Google One subscriber for the extra storage, AI was being bundled in with that package) to make some rote changes to finish building the RSVP form. I was blown away by how effective it was. It took half the time that it would have taken me to add and style the fields, the validation logic and the unit tests. This did not help the spiral and insecurities I had about returning to work. ## Professionally But back to work I went, and even though I don’t write much code any more, I have continued to use AI tools to assist me when I do, as well as at home on my projects. We now have access to Claude Code at work (previously we only had Gemini in a web interface) and the level of integration it offers is excellent when it comes to working through and debugging problems, and through using it you get to understand the quirks, how to guide it to get the results you want, and also get annoyed at it when it goes completely off-piste (for example, when the linter failed on a task I had recently, it just added lots of `eslint-disable` s, rather than fixing the actual flagged issue). I like Claude, but at home I now use JetBrains AI Assistant after Google Code Assistant was discontinued (just the free tier, as I use it occasionally enough to not need more than that) and I am finding it has a superior integration into IntelliJ (my IDE of choice). Gemini is also an excellent tool for debugging, but I guess it’s to be expected that this is a task LLMs do well. With the sheer volume of training material it has, plus the ability to look things up in Google’s index, identifying crucial log lines and matching that to discussions in GitHub issues and other sources to summarise the problem and a solution sounds like a task it is well suited for. At times it also goes wildly wrong at diagnosing the issue – sometimes it misses crucial context (and it doesn’t know to ask for it) and sends me down a rabbithole, or it doesn’t realise it’s working on an outdated solution. But on balance it’s right more often than it’s wrong, and even when it’s wrong, it’s a mildly frustrating [rubber duck](https://en.wikipedia.org/wiki/Rubber_duck_debugging). It is annoying how unreasonably effective LLMs are, and I am not the first software engineer to discover it. As tech lead, I try to encourage the developers in the teams I look after to give it a go. I have no doubt that it’ll help them out. But I also know the ethical tradeoffs I’ve made to accept it aren’t for everyone, so I don’t want to force it upon them, in the same way that while I believe JetBrains IDEs are far less frustrating to use than VS Code, but I’m not going to force people to change their tool of choice. But I’ve been asked what happens if someone chooses to opt out of LLM usage for ethical reasons, but fall behind the performance of their colleagues, and I struggle to answer. I want to say it doesn’t matter, but I know that’s not necessarily how it’ll be seen organisationally. ## Politically I’ve also started using LLMs in my political career too. But unlike Labour who brought forward an empty motion on district centres that reeked of AI tells ([which I called out at the time](https://www.manchestereveningnews.co.uk/news/greater-manchester-news/plan-make-harpurhey-next-chorlton-31237708)), or the Green leader who stood up to speak at the last full Council meeting and started with [the AI tell of “I rise to speak”](https://www.yahoo.com/news/articles/chatgpt-triggers-surge-mps-using-132959577.html?guccounter=1), I’m trying to use it to help me be better, rather than replace my political skills. I read every paper that comes to the committee I come to, but these papers often come with appendices that can be very detailed and deep, and often is where the real detail hides. It’s easy to miss things when reviewing them. But doing what I once saw whilst sat next to a Labour councillor, which was to copy and paste a paper into ChatGPT, type “what would be good questions to ask” and then ask those questions without taking the time to understand them is not right. I still read the committee papers and draw up my first round of key points in my notebook. But then I copy the papers and the appendices into Gemini and start asking it some surface deep questions that I might previously have gone to an officer first on, allowing me to make sure I’ve not missed anything and refine my questions down to the most effective ones I want to get to in scrutiny. While, of course, sense checking anything that’s flagged by the LLM back to the source document – it does often confuse two nearby concepts especially around financial spending. Again, this is all leaning on the exceptional ability of LLMs in text summarisation, and I take care to not outsource my thinking (I was elected, not Gemini), but to instead help me. I do sometimes use the text generation function of LLMs, especially for generating form letters to third parties where I’m helping advocate on behalf of residents (the party has reviewed and approved Gemini for use on paid party accounts). I do usually edit the first draft and make sure that I’m hitting the right tone (usually constructive to start with) rather than fire and forget what it said. There is a flipside of this though. Residents are also using LLMs to generate casework emails to me. I can usually tell straight away – most emails I get are for people asking for help, often in crisis, but sometimes just struggling to navigate the council’s services. But there’s been a surge in emails which are not asking for help, they’re demanding I tackle their problem by taking particular action, often quite forcefully, or disproportionally, such as wanting to kick off a full investigation for a mattress at the side of the road (something I’d jump at doing if I thought there was any chance of catching the perpetrator). This doesn’t fit with my casework approach and actually although they’re clearly trying to get a good outcome, might actually be counter-productive for them, as it can set up a different relationship between us than just helping someone in trouble. I’ve developed my own relationships with council officers and navigating the internal politics and bureaucracy of the council to get things done is an important skill which often results in a different approach than the list of demands from ChatGPT expects. It’s definitely made casework harder overall. I’ve also found myself using AI to review things I write. At this week’s Lib Dem Conference I submitted a motion that I asked Gemini to give a pass over. It surprised me with the understanding it had of the party’s internal processes, and started using internal jargon like “[FCC](https://en.wikipedia.org/wiki/Liberal_Democrat_Conference#Federal_Conference_Committee)” when giving me feedback, which I did agree with and incorporated into the final draft. Ultimately that amendment wasn’t selected though (despite the reassurance Gemini gave me it was bulletproof!) so don’t take this as a particularly glowing review. ## Personally But I think the thing that’s surprised me about my AI usage is how I’m now using it in my personal life. Just before Manchester Pride, I decided to get lash extensions as an experiment. I love false lashes when I’ve worn them on nights out, and my go to makeup look was usually just mascara. I absolutely loved how I looked, and it unlocked something in me. It also made me face up to some internal unhappiness with my current style and wardrobe, which I didn’t feel was fully expressive of me. With coming up to 5 years on HRT, I’ve seen not only body shape changes in that time that my clothes didn’t necessarily flatter, but also significant weight loss due to Wegovy (I’ve dropped 25kg now, going to a size 14 from a size 18). So I’ve toyed with the idea of tinted moisturiser for a while to add a bit more of a glow to my daily look, and off the back of the lash extension inspired motivation, Meg and I went to Sephora and asked for some advice. We were given a product recommendation, which I bought and took home, but I struggled to have the confidence to know how to effectively use it. When I got home, I popped the new products into Gemini, alongside my existing skincare routine and asked it to help me build out a 5 minute morning routine I could use to develop the habit in my morning routine. Things escalated from there. I started asking it to evaluate my hair routine, and when it came to the hen do, I asked it for how to update my full face makeup routine and came up with a list of sensible things to try, that resulted in me spending quite a bit more at Sephora… These are all things I’ve wanted to do for years, but I’ve struggled to know how to even get started. I find makeup communities intimidating and get a real imposter syndrome in them and don’t even know how to start navigating the basics. The only real skills I have were picked up from a masterclass ran by [Patti](https://www.wowgals.co.uk/about) at Manchester Pride a few years ago but this was a full glam look – a bit much for every day, both in terms of the end result and the time taken to apply. But I found Gemini non-judgemental, knowledgeable and happy to explain things in detail. I continued down the rabbit hole and started describing my style and my aesthetic goals to give the prompt a reference, and before I knew it I had a reference doc I could refer to, and then re-use for future prompts. I started looking at refreshing my wardrobe, and asked Gemini for the right search terms for ASOS to spend a gift voucher I got for my birthday for things that fit my goals. I told it when I didn’t like what I found, and it refined it. When it summarised what I asked for it introduced me to terms I didn’t know (turns out I had no idea what a capsule wardrobe but was despite hearing the term before, but that’s what I wanted!), and when I asked it to cross check my goals and my action against each other it even came up with a style summary for me that accurately reflects what I’m going for with my personal expression, style and aesthetics: *“doll who definitely has a GitHub account”*. ## Losing Myself I think I now understand how people fall into AI psychosis. My fiancée has certainly expressed some concern that I’ve gone a bit far with it, so I’ve had to make sure I check myself. Unlike in my professional and political life where I have the knowledge to be critical of AI output, I definitely don’t when it comes to fashion and beauty and had to accept what it said, which was all very plausible (as text generation tends to be) but I had no way of evaluating if it was right. I tried to correct this in the AI itself, asking it to be critical or brutal when assessing my goals and the plan against it which would pull up contradictions, but after I bought my capsule wardrobe to start wearing, I realised I’d lost my self-expression within this, my personal geeky side. So now I’m bringing it back in, yet I’m still using the AI to help me do this. I need to make sure I don’t just trust it uncritically and lose myself in it. I have to talk to people too, whether that’s my partner, or my friends (including the one—I have to remind myself—who used to be a fashion journalist). And I think that’s one of my biggest concerns about this tool, which is the risk of losing community. Although I never felt confident to do this and that was a barrier for me actually doing this, if I’d wanted to understand myself better and understand style and beauty better, I’d have had to throw myself into something like a sub-Reddit to start discussing this, learning from other people and synthesising the information myself. But if I’d done that, I’d have also been posting back, communicating and maybe making some friends along the way. Yet I’ve not done that (although I’ve talked to my existing friends group about what I’m doing), and I worry what I’ve lost, and what we’re all losing at scale. Writing this blog post is a way of giving back, and I’m happy to share details of my routines and products with people, but I’m worried about posting it with no context as what works for me and my style goals will not for other people, and I don’t know how to give the context to evaluate it. ## My ethical limits And so, I can’t ignore the ethics. I will be honest, I do at times feel hypocritical about this, in the same way that I feel discomfort in knowing I eat meat despite the horrors of the factory farming industry. It does at times feel like a moral failing to be using this technology given the environmental impact of the roll out of LLMs by the hyperscalers (or “[blitz-scaling” as a recent paper to the committee I sit on in Manchester](https://democracy.manchester.gov.uk/documents/s65338/Appendix%201%20for%20AI%20and%20Business%20Adoption%20in%20Manchester%20Opportunities%20Risks%20Labour%20Market%20Change%20an.pdf) called it), as well as the mania of investment that surrounds it feeling unhealthy. But I also strongly believe that knowledge should be free and shared (I wrote my book after all not to make a living, but to share my skills), and drawing a line which excludes the use of that knowledge by AI firms doesn’t sit well with me. One of the driving factors when I wrote *[The Full Stack Developer](https://link.amazon/B0imXx6cv)* was to share my knowledge; I never expected to make money off of it. And in tech I’m a big believer in open source, and coding in the open, and it’s a genuine benefit of my profession that we have a strong culture of sharing. That said, although I believe knowledge and information should be open and shared, I think creative expression is personal and to be protected, and the use of AI to generate “slop” prose, images or video undermines creatives who make their living off their work, and is genuinely polluting (and risks model collapse) and harmful. LLMs are currently both, and it’s a fine line to balance. I’ll ask Gemini to give me feedback on this blog post, but I’m not going to ask it to generate it for me from my bullet points. When I studied my MSc, machine learning made up a big part of it, and although the techniques and technologies are unrecognisable from the state of the art in 2010 when I graduated, I can see where it’s come from, and I’ve built a mental model of LLMs around what I studied back then. I believe LLMs are excellent at tackling those AI problem areas such as summarisation, and despite Noam Chomsky’s claim that the idea of the probability of a sentence is “useless”, are also very good at generating meaningful text given a well enough structured prompt and enough context, despite being a [stochastic parrot](https://en.wikipedia.org/wiki/Stochastic_parrot). But I also know that although a probable text can be useful, there’s no guarantee it’s right, and I believe consciousness is more than just chaining tokens. Agents are also incredibly powerful, with LLM models providing control to these, but they’re also dangerous. Last time I saw my former BBC R&D colleague [Rosie Campbell](https://www.rosiecampbell.xyz/) (who was so alarmed by AI risk she moved to the US and shifted her career to focus on this risk) I was sceptical that the level of risk these agents posed were at the same level as her fears. But now seeing how quickly the technology has progressed, and seeing agents in action, I now regret the short-sightedness about the risk and can see scenarios where either unchecked agents, or maliciously tasked agents can cause global-scale challenges. But despite all that, I’m still using LLMs. Just this week I got Claude to diagnose a long-running bug at work (some login paths incorrectly returning a 500 rather than a 401) which ended up being in an upstream dependency and I’ve been using it to help refine my search for shoes as part of my wardrobe refresh. Am I part of the problem? Maybe. But it’s unlocked a level of my transition I was struggling with, so right now it’s hard to feel anything other than overall positive about the most recent changes in my life which AI has helped enable.