Prompting with Context - My Friend's Fish & My Tinder Bio

My first attempt at Tinder, fourteen years ago, had only one photo: me, standing beside a fish my friend had caught, holding it like it was mine. No bio, no context. Nothing for anyone to actually go on, including the fairly basic fact of whether I'd caught the fish myself or was simply present when someone else did. I got no matches, not one. A few weeks later I added an actual bio and swapped the fish for a photo of myself with my dog, and suddenly I had two matches. Both turned out to be bots trying to con me out of money, but I'm still counting it as a win, because it proved the point. Same fella, same face mostly, completely different result, purely because I'd finally given people something to go on. It took me an embarrassingly long time to realise the exact same thing was happening every time I got a useless answer out of ChatGPT. Last year I needed to get my head around French einvoicing. New mandate, plenty of moving parts, a deadline I hadn't fully looked at yet, and the general sense of dread that comes with any sentence containing the words "Chorus Pro. I'm using it here because it's the clearest example I have of what prompting with context" actually means in practice, rather than as a phrase people throw around in LinkedIn posts. So I did the obvious thing and typed "tell me about e-invoicing in France" into ChatGPT, fully expecting it to do my job for me. It did not do my job for me. What I got back read like the kind of answer you give in a meeting when you haven't done the reading but don't want anyone to know. Technically about the right topic. Confident but useless. I'd asked it nothing, really, and it had nothing to give me, and yet it gave it to me anyway, at length, with great conviction. I went back and forth with it for what was an embarrassingly long time before I worked out the problem wasn't the model. It was that I'd handed it a sticky note that said "sort out France" and walked off, the same way I normally brief a new team member. So I started again, properly, and told it five things. What it should pretend to be I asked it to answer as an einvoicing specialist from a Big4 firm. A specific, expert who's spent years reading DGFiP guidance and has opinions on the difference between Chorus Pro and a PDP. The moment I gave it that costume, it stopped sounding like an encyclopedia and started sounding like someone who actually does this for a living. What our situation actually is Then the facts, properly laid out rather than implied. We're a B2B software company, headquartered in the US. French B2B customers are serviced through a local selling entity there. B2C sales irun through an IOSS return filed out of Ireland. And, the bit that turned out to matter most, we have an actual office in France, with employees, local AP, and a steady trickle of expenses going through French books. Once it had that, the answer narrowed from "here's the mandate" to "here's what a locally established business with French AP specifically needs to do," which is a smaller and far more useful problem. Who I am I told it I'm an accountant with a solid VAT background, so it could skip the bit where it explains the very basics, but to reply in plain business language rather than a memo written to impress another tax adviser. I don't need proof of homework. I need something I can act on or hand to a colleague without translating it first. Where it's getting this from I told it to cite a source for every claim, and to keep those sources to government sites, recognised consulting firms, or anything I specified myself. No more confident, unattributable statements I'd have to go and verify from scratch anyway, which rather defeats the point of asking in the first place. That I don't trust it And finally, I told it plainly: I don't trust you. Treble check everything before you hand it to me, and expect to be grilled on every claim once you do. It's an odd thing to say to a piece of software, but it worked. The tone shifted. Less breezy certainty, more flagging of where the guidance was genuinely unclear or where I should check before relying on it. Which is exactly what I'd want from a person giving me tax advice, so I'm not sure why I expected less from the software. Lastly Same model, same original question, an entirely different and far more useful answer. The first attempt was vague enough to be worthless. The second was specific enough to be a genuine starting point, the kind of first draft you'd actually be glad to get from a new associate, fully sourced, and fully aware I was going to check every line of it anyway. None of it was clever. Tell it who to be, what your situation actually is, who you are, where its claims come from, and that you don't believe a word of it until it's proven. It's not a trick. It's just briefing someone properly instead of leaving them to guess, which is advice I probably should have been giving my actual junior staff for years. I hope you found this helpful Colm