WEBVTT NOTE Sentence-level transcript of https://www.youtube.com/watch?v=gpmze-6ueGU NOTE One cue per sentence. Cue ids are the line anchors on /transcripts/gpmze-6ueGU.html. A cue ends where the next begins, or 2 s after its last word. s1 00:00:01.600 --> 00:00:05.839 Hi everyone, I'm Matt Rigon and I'm a financial planner at Range. s2 00:00:05.839 --> 00:00:16.640 I want to show you how our financial planners ended up writing the agents that we use every day and none of us the financial planners for that matter write code. s3 00:00:16.640 --> 00:00:18.320 This is the build guide. s4 00:00:18.320 --> 00:00:28.960 How we did it step by step so you can take it back to your own teams, companies, etc. So, a quick bit about me because it matters for how you'll hear this talk. s5 00:00:28.960 --> 00:00:30.240 I'm not an engineer. s6 00:00:30.240 --> 00:00:38.800 I'm a certified financial planner and I spend my days with clients talking about cash flow, taxes, insurance, retirement, estate, you name it. s7 00:00:38.800 --> 00:00:46.559 I also lead the AI adoption for the planning team at Range, which is an all-in-one wealth management company. s8 00:00:46.559 --> 00:00:51.680 Most agent talks you'll hear come from the people who built the platform. s9 00:00:51.680 --> 00:00:56.719 This talk comes from someone on the other side, the people you hand the platform to. s10 00:00:56.719 --> 00:00:58.559 We work in a regulated business. s11 00:00:58.559 --> 00:01:02.719 The output is financial advice to real clients. s12 00:01:02.719 --> 00:01:05.360 And the review step is not optional. s13 00:01:05.360 --> 00:01:07.360 So, this isn't a vendor pitch. s14 00:01:07.360 --> 00:01:12.320 It's a method and it's what actually worked for us. s15 00:01:12.320 --> 00:01:16.240 We started where everyone starts AI as a chatbot. s16 00:01:16.240 --> 00:01:21.520 I'd write a client message, paste it in, and ask it to make it a little clearer or friendlier. s17 00:01:21.520 --> 00:01:22.479 And that was useful. s18 00:01:22.479 --> 00:01:28.080 It saved me a few minutes here and there, but I was still doing all of the work manually. s19 00:01:28.080 --> 00:01:30.880 I was just doing it a little bit faster. s20 00:01:30.880 --> 00:01:35.040 The change that actually mattered came later. s21 00:01:35.040 --> 00:01:36.560 Here's what we figured out. s22 00:01:36.560 --> 00:01:40.960 A planner's constraint was never typing speed, for example. s23 00:01:40.960 --> 00:01:44.560 It's that we can only be in one place at a time. s24 00:01:44.560 --> 00:01:49.119 If I'm on a call with a client, I'm not building someone else's plan. s25 00:01:49.119 --> 00:01:51.840 Writing faster doesn't fix that problem. s26 00:01:51.840 --> 00:01:56.159 What fixes it is work that happens while I'm somewhere else. s27 00:01:56.159 --> 00:01:59.200 So, here's what a day looks like now. s28 00:01:59.200 --> 00:02:01.360 I'm on a call with one client. s29 00:02:01.360 --> 00:02:09.679 At the same time, an agent is drafting a full cash flow analysis or an insurance analysis for a different client. s30 00:02:09.679 --> 00:02:12.160 When the call ends, I open that draft. s31 00:02:12.160 --> 00:02:18.319 I fully review it, adjust what needs to be adjusted, and send the client their recommendations. s32 00:02:18.319 --> 00:02:21.120 I want to be really clear about that last step. s33 00:02:21.120 --> 00:02:23.760 The agent doesn't send anything. s34 00:02:23.760 --> 00:02:29.599 A planner reviews every single recommendation before a client sees it. s35 00:02:29.599 --> 00:02:35.599 What changed is that the first draft no longer needs me sitting in front of it. s36 00:02:35.599 --> 00:02:38.480 Getting there took a lot more than a good prompt. s37 00:02:38.480 --> 00:02:41.120 However, two things had to be true. s38 00:02:41.120 --> 00:02:49.360 First, the agent has to work where the work lives, not in a separate chat window I'm copying and pasting it into. s39 00:02:49.360 --> 00:02:54.879 Second, it has to know how we do it, not how financial planning works in general. s40 00:02:54.879 --> 00:03:00.560 The model already knows that, but the agent needs to know how range does it. s41 00:03:00.560 --> 00:03:05.760 What we look at first, what we always flag, how we phrase a specific recommendation. s42 00:03:05.760 --> 00:03:08.319 That combination is what a skill is. s43 00:03:08.319 --> 00:03:16.800 Our firm's way of doing one specific piece of work written down so an agent can follow it. s44 00:03:16.879 --> 00:03:21.440 The bulk of this talk is how you build one. s45 00:03:21.440 --> 00:03:22.640 Five steps. s46 00:03:22.640 --> 00:03:26.879 Pick, teach, pack, test, and then ship. s47 00:03:26.879 --> 00:03:29.280 None of them require writing code. s48 00:03:29.280 --> 00:03:39.840 Every one of them requires knowing the work really well, which is exactly why the people doing the work should also be the ones building it. s49 00:03:40.239 --> 00:03:43.120 Step one, pick a workflow you repeat. s50 00:03:43.120 --> 00:03:49.840 Your initial instinct is to go for the flashiest work, whatever's taking the most amount of time and the most complex. s51 00:03:49.840 --> 00:03:58.319 Don't pick something you do every week where you already know what good looks like, where you can spot a bad draft at a glance. s52 00:03:58.319 --> 00:04:07.120 Additionally, we recommend choosing some of the least complex workflows to start, but some that require the most manual work. s53 00:04:07.120 --> 00:04:12.560 For us, that meant things like cash flow analysis and insurance reviews. s54 00:04:12.560 --> 00:04:16.560 Just for us here at range, that required the most manual work. s55 00:04:16.560 --> 00:04:21.759 And if you can't judge the output quickly, you can't test it and you can't trust it. s56 00:04:21.759 --> 00:04:23.440 So look for two things. s57 00:04:23.440 --> 00:04:28.160 You do it often and you know how to judge it. s58 00:04:28.320 --> 00:04:33.600 Step two, show it several real cases for previous plans you've delivered. s59 00:04:33.600 --> 00:04:39.360 Take several actual client cases and walk through the work start to finish. s60 00:04:39.360 --> 00:04:42.639 And as you do it, narrate your judgment out loud. s61 00:04:42.639 --> 00:04:51.360 Not just next I open the pastub, why you're opening it, what you're looking for, what would make you stop and ask a question. s62 00:04:51.360 --> 00:04:54.080 That narration is the whole game. s63 00:04:54.080 --> 00:04:55.759 The steps are easy to write down. s64 00:04:55.759 --> 00:05:03.040 The judgment is what makes it your firm's work instead of generic advice. s65 00:05:03.600 --> 00:05:06.160 Step three, pack it into a skill. s66 00:05:06.160 --> 00:05:09.280 A skill isn't just a lip a list of steps. s67 00:05:09.280 --> 00:05:10.960 Two things matter most. s68 00:05:10.960 --> 00:05:14.080 The questions it must ask before it starts. s69 00:05:14.080 --> 00:05:15.680 What does it need to know? s70 00:05:15.680 --> 00:05:19.039 What should it confirm with me instead of simply guessing? s71 00:05:19.039 --> 00:05:21.680 And the limits it must respect. s72 00:05:21.680 --> 00:05:23.840 What it never does on its own. s73 00:05:23.840 --> 00:05:28.560 In our world, the big one is that it never sends anything to a client. s74 00:05:28.560 --> 00:05:33.919 It drafts and then a planner decides if it gets sent or needs to be adjusted. s75 00:05:33.919 --> 00:05:38.720 If you only write the steps, you get something that simply just runs on its own. s76 00:05:38.720 --> 00:05:43.520 If you write the questions and set the limits, you can get something you can trust. s77 00:05:43.520 --> 00:05:45.759 Think of a skill as a recipe. s78 00:05:45.759 --> 00:05:51.919 Once Claude knows how to do the skill, it will repeat your process on command. s79 00:05:51.919 --> 00:05:59.199 Step number four, run it five or 10 times on real cases before anyone else sees it. s80 00:05:59.199 --> 00:06:02.000 This is the step everyone wants to skip. s81 00:06:02.000 --> 00:06:05.120 Every run teaches you something the skill didn't know yet. s82 00:06:05.120 --> 00:06:10.560 An edge case, a question it should have asked, a spot where it sounded confident and was wrong. s83 00:06:10.560 --> 00:06:16.639 You need to fix it, save the skill, and run it again and again and again. s84 00:06:16.639 --> 00:06:28.080 Five to 10 runs sounds like a lot, but it's a lot less work than cleaning up after you hand a halfbaked skill to 15, 20, 100 people in your team. s85 00:06:28.080 --> 00:06:32.560 Step five, ship it with an owner and a reviewer. s86 00:06:32.560 --> 00:06:34.720 The owner keeps it current. s87 00:06:34.720 --> 00:06:41.840 Our processes change, tax rules change, and a skill that was right in March can be wrong by December. s88 00:06:41.840 --> 00:06:45.199 The reviewer checks it before the team relies on it. s89 00:06:45.199 --> 00:06:48.400 And it's someone other than the author. s90 00:06:48.400 --> 00:06:51.440 Same as a plan, you don't grade your own work. s91 00:06:51.440 --> 00:06:56.160 Once other people depend on a skill, somebody has to be accountable for it. s92 00:06:56.160 --> 00:07:03.199 That's the difference between a clever personal shortcut and something a team can build on. s93 00:07:03.199 --> 00:07:05.199 So, where are we at currently? s94 00:07:05.199 --> 00:07:11.199 Six months in, we have 24 skills across the company and 17 of them are in planning. s95 00:07:11.199 --> 00:07:13.520 And here's the part I didn't expect. s96 00:07:13.520 --> 00:07:18.000 It spread to sales, product, marketing, all without a mandate. s97 00:07:18.000 --> 00:07:19.520 Nobody sent a memo. s98 00:07:19.520 --> 00:07:24.639 People saw what the planners were doing and started building their own agents. s99 00:07:24.639 --> 00:07:33.919 Every one of those skills was written and is maintained by planners, marketers, and salespeople, not developers or engineers. s100 00:07:34.160 --> 00:07:40.639 A couple of things changed once subject matter experts started building for their own niche. s101 00:07:40.639 --> 00:07:42.960 The skills got a lot more specific. s102 00:07:42.960 --> 00:07:47.599 The person who knows insurance, for example, best builds the insurance skill. s103 00:07:47.599 --> 00:07:53.280 The person who best knows tax, tax code, tax law should be building the tax one. s104 00:07:53.280 --> 00:07:55.280 and a review came naturally. s105 00:07:55.280 --> 00:07:59.520 We already review each other's plans and that's just how planning works. s106 00:07:59.520 --> 00:08:03.440 So now we review each other's skills in that same way. s107 00:08:03.440 --> 00:08:08.160 It's the same habit, just a new thing to review. s108 00:08:08.240 --> 00:08:10.960 So why does this matter right now? s109 00:08:10.960 --> 00:08:13.360 Because the bottleneck has moved. s110 00:08:13.360 --> 00:08:17.919 Whether a model can draft a cash flow analysis, that question's already been settled. s111 00:08:17.919 --> 00:08:18.879 It can. s112 00:08:18.879 --> 00:08:20.400 What's not settled? s113 00:08:20.400 --> 00:08:22.800 Who writes down how your firm does it? s114 00:08:22.800 --> 00:08:28.560 who tests it before it ships, who owns it once 15 other people depend on it. s115 00:08:28.560 --> 00:08:35.839 If you're being asked to push agents out past the engineering org, you're about to hit these exact questions. s116 00:08:35.839 --> 00:08:42.159 There isn't much field data on this yet, but this is how we've done it here at Range. s117 00:08:43.200 --> 00:08:44.159 So, that's it. s118 00:08:44.159 --> 00:08:46.000 Thank you so much for listening to this talk. s119 00:08:46.000 --> 00:08:50.720 I'm Matt Rigon from Range and you can reach me for at mattrange.com. s120 00:08:50.720 --> 00:08:53.200 Have a great day.