WEBVTT Kind: captions Language: en NOTE Module 1: Foundations narrated walkthrough. Generated from the narration script and per-segment audio durations (0.7 s gap per slide). s01-001 00:00:00.000 --> 00:00:02.493 Welcome to Claude Code for Accounting Research. s01-002 00:00:02.543 --> 00:00:06.285 I'm John Barrios, and this is Module 1, Foundations: tools, context, s01-003 00:00:06.335 --> 00:00:07.893 and the research map. s01-004 00:00:07.943 --> 00:00:10.521 So today we're going to get the tool working, s01-005 00:00:10.571 --> 00:00:15.310 and then build the habits that make it something you can trust with research work. s02-006 00:00:16.060 --> 00:00:18.297 So here's how the live session runs. s02-007 00:00:18.347 --> 00:00:22.907 We start with about forty minutes of lecture on tools, context, verification, s02-008 00:00:22.957 --> 00:00:24.460 and the research map. s02-009 00:00:24.510 --> 00:00:29.072 Then a thirty-five minute live demo, where I take one task from start to finish. s02-010 00:00:29.122 --> 00:00:30.871 A ten-minute break. s02-011 00:00:30.921 --> 00:00:34.109 Then eighty minutes of lab, where you install Claude Code, write a CLAUDE.md file, s02-012 00:00:34.159 --> 00:00:37.967 and run one verified task. s02-013 00:00:38.017 --> 00:00:40.882 And we close with a debrief and an exit ticket. s02-014 00:00:40.932 --> 00:00:42.377 This video walks through the lecture slides, s02-015 00:00:42.427 --> 00:00:46.650 so you get a preview of the demo and the lab here, not the full thing. s03-016 00:00:47.400 --> 00:00:49.208 You've probably heard this one... s03-017 00:00:49.258 --> 00:00:51.163 or maybe you've said it. s03-018 00:00:51.213 --> 00:00:55.452 You ask a chat tool to open a Compustat file on your desktop and plot median s03-019 00:00:55.502 --> 00:00:59.542 return on assets by industry, and it tells you it can't see your files. s03-020 00:00:59.592 --> 00:01:01.891 And that's true, it can't. s03-021 00:01:01.941 --> 00:01:03.295 Now look at the right side. s03-022 00:01:03.345 --> 00:01:05.363 Same request in Claude Code. s03-023 00:01:05.413 --> 00:01:07.496 It reads the file, writes a script, runs it, s03-024 00:01:07.546 --> 00:01:11.256 and hands back one number for you to check by hand. s03-025 00:01:11.306 --> 00:01:14.368 Both panels are stylized, they're not screenshots. s03-026 00:01:14.418 --> 00:01:17.732 So when people say AI can't do real research work, s03-027 00:01:17.782 --> 00:01:21.388 a lot of times what they mean is the tool they tried can't do it. s03-028 00:01:21.438 --> 00:01:24.630 Different tools can do different things, and that's the next slide. s04-029 00:01:25.380 --> 00:01:28.321 So think of this as a ladder with five rungs. s04-030 00:01:28.371 --> 00:01:30.079 At the bottom is chat. s04-031 00:01:30.129 --> 00:01:35.208 That's fine for a definition or a syntax question, but it can't work with real data. s04-032 00:01:35.258 --> 00:01:39.700 Next is an inline assistant in your editor that autocompletes, say, s04-033 00:01:39.750 --> 00:01:40.673 a regression script. s04-034 00:01:40.723 --> 00:01:45.716 Then an agentic IDE, which can make a chain of edits across a few files. s04-035 00:01:45.766 --> 00:01:47.776 Claude Code is on rung four. s04-036 00:01:47.826 --> 00:01:52.108 It works in your terminal, reads your files, and runs your code end to end. s04-037 00:01:52.158 --> 00:01:57.296 Rung five connects the agent to outside services, like querying WRDS directly, s04-038 00:01:57.346 --> 00:01:59.289 and we do that in Module 4. s04-039 00:01:59.339 --> 00:02:02.770 And Cowork, by default, sits at or below rung two. s05-040 00:02:03.520 --> 00:02:05.062 Here's a common failure. s05-041 00:02:05.112 --> 00:02:08.633 You tried a browser assistant, it couldn't open your data file, s05-042 00:02:08.683 --> 00:02:11.619 and you concluded AI can't help with research. s05-043 00:02:11.669 --> 00:02:13.915 That's a reasonable experience to have had, s05-044 00:02:13.965 --> 00:02:17.385 and a lot of students come into this course after exactly that. s05-045 00:02:17.435 --> 00:02:19.702 But the tool was doing what it was built to do, s05-046 00:02:19.752 --> 00:02:22.032 and it wasn't built to touch your filesystem. s05-047 00:02:22.082 --> 00:02:24.848 So ask which rung the task needs. s05-048 00:02:24.898 --> 00:02:29.034 If it has to touch real files, run real code, and iterate on its own, s05-049 00:02:29.084 --> 00:02:30.830 you need rung four or above. s06-050 00:02:31.580 --> 00:02:34.477 This is what one Claude Code turn looks like. s06-051 00:02:34.527 --> 00:02:36.393 Number one is your prompt. s06-052 00:02:36.443 --> 00:02:40.861 Number two is the tool calls, so that's the agent acting on your machine: s06-053 00:02:40.911 --> 00:02:43.897 it reads the file, writes the script, runs it. s06-054 00:02:43.947 --> 00:02:47.135 Number three is what it saw, and you can read all of it. s06-055 00:02:47.185 --> 00:02:51.269 Four is the handback, ideally with one number for you to verify. s06-056 00:02:51.319 --> 00:02:53.122 And five tells you where you are: which folder, which model, s06-057 00:02:53.172 --> 00:02:56.625 and how much context you've used. s06-058 00:02:56.675 --> 00:02:59.879 So think of it as a research assistant sitting at your terminal. s06-059 00:02:59.929 --> 00:03:04.810 It uses your own R, Python, or Stata, and you manage it the way you'd manage an RA. s07-060 00:03:05.560 --> 00:03:07.127 A related failure. s07-061 00:03:07.177 --> 00:03:11.217 Because it feels conversational, people assume it comes with its own Python, s07-062 00:03:11.267 --> 00:03:14.047 the way a cloud notebook does. It doesn't. s07-063 00:03:14.097 --> 00:03:19.544 It has no Python or R of its own, so it uses whatever's installed on your machine. s07-064 00:03:19.594 --> 00:03:22.531 If a script it writes fails with command not found, s07-065 00:03:22.581 --> 00:03:25.297 check your environment before you blame the agent. s07-066 00:03:25.347 --> 00:03:27.246 Windows is usually where this shows up, s07-067 00:03:27.296 --> 00:03:31.374 because a native Python install and one inside WSL can be different, s07-068 00:03:31.424 --> 00:03:32.710 and nothing warns you. s08-069 00:03:33.460 --> 00:03:36.022 Now, how the context window works. s08-070 00:03:36.072 --> 00:03:39.320 Claude Code doesn't remember your project between turns. s08-071 00:03:39.370 --> 00:03:40.736 Every time you send a message, s08-072 00:03:40.786 --> 00:03:45.682 the whole conversation gets sent again as one package: every prompt, every reply, s08-073 00:03:45.732 --> 00:03:47.885 every file it read in. s08-074 00:03:47.935 --> 00:03:50.138 The standard window is two hundred thousand tokens, s08-075 00:03:50.188 --> 00:03:52.801 and some models go up to a million. s08-076 00:03:52.851 --> 00:03:54.388 That sounds like a lot, right? s08-077 00:03:54.438 --> 00:03:58.991 But one large CSV read straight into the conversation can use a sizable share of s08-078 00:03:59.041 --> 00:04:00.473 it in one turn. s08-079 00:04:00.523 --> 00:04:04.515 So don't have it read a giant Compustat extract just to look at it. s08-080 00:04:04.565 --> 00:04:08.864 Ask for a script that processes the file on disk and reports back a summary. s08-081 00:04:08.914 --> 00:04:12.850 And you can type /context at any point to see how full the window is. s09-082 00:04:13.600 --> 00:04:16.193 This next one is hard to unlearn. s09-083 00:04:16.243 --> 00:04:19.183 People treat a session like a shared drive or an email thread, s09-084 00:04:19.233 --> 00:04:22.842 where last week's work is still there because it's in the account. s09-085 00:04:22.892 --> 00:04:24.955 A session doesn't work that way. s09-086 00:04:25.005 --> 00:04:29.501 Nothing carries over from one session to the next unless it was written to a s09-087 00:04:29.551 --> 00:04:31.285 file a future session can read. s09-088 00:04:31.335 --> 00:04:34.265 That's the main change in how you think about this tool. s09-089 00:04:34.315 --> 00:04:37.577 CLAUDE.md, progress files, named scripts... s09-090 00:04:37.627 --> 00:04:40.590 they're all there because sessions don't keep any memory. s10-091 00:04:41.340 --> 00:04:44.100 As a session runs long, two things happen. s10-092 00:04:44.150 --> 00:04:48.271 The window fills up, and quality can drop well before it's full, s10-093 00:04:48.321 --> 00:04:51.968 because all the trial and error and dead ends get re-sent every turn, s10-094 00:04:52.018 --> 00:04:57.188 and that noise makes the model reason less carefully. s10-095 00:04:57.238 --> 00:04:58.019 So you compact on purpose. s10-096 00:04:58.069 --> 00:05:02.498 First, ask Claude to write a short progress.md: what's been decided, s10-097 00:05:02.548 --> 00:05:04.331 which files exist, what's left. s10-098 00:05:04.381 --> 00:05:09.060 Then run /compact, which condenses the history into a short summary. s10-099 00:05:09.110 --> 00:05:13.200 For a clean start, open a fresh session and have it read that file. s10-100 00:05:13.250 --> 00:05:15.978 And keep CLAUDE.md itself short. s10-101 00:05:16.028 --> 00:05:19.055 Anthropic's guidance is under about two hundred lines. s10-102 00:05:19.105 --> 00:05:23.311 Facts go in CLAUDE.md, and step-by-step procedures go in skills, s10-103 00:05:23.361 --> 00:05:25.150 which you'll see in Module 3. s10-104 00:05:25.200 --> 00:05:29.370 That's mantra one on the slide: files persist, context doesn't. s11-105 00:05:30.120 --> 00:05:35.062 Compaction replaces a long, noisy transcript with a short summary of what mattered, s11-106 00:05:35.112 --> 00:05:38.754 the way you'd hand a project to a colleague with a two-paragraph memo. s11-107 00:05:38.804 --> 00:05:40.840 People get this wrong in two ways. s11-108 00:05:40.890 --> 00:05:45.164 Some avoid /compact because they're afraid of losing progress. s11-109 00:05:45.214 --> 00:05:49.357 Others trust it without writing the important decisions to a file first. s11-110 00:05:49.407 --> 00:05:54.044 If something specific has to survive, tell Claude to keep it when it compacts. s11-111 00:05:54.094 --> 00:05:56.950 And put those decisions in a file before you run it. s12-112 00:05:57.700 --> 00:06:00.233 If Claude Code is an RA at your terminal, s12-113 00:06:00.283 --> 00:06:03.265 you treat its work the way you'd treat a real RA's work. s12-114 00:06:03.315 --> 00:06:06.651 You delegate, and you check the result before you build on it. s12-115 00:06:06.701 --> 00:06:10.320 So say it builds a leverage panel from a Compustat extract. s12-116 00:06:10.370 --> 00:06:12.442 Don't take the whole panel on faith. s12-117 00:06:12.492 --> 00:06:14.095 Pick one firm you know well, s12-118 00:06:14.145 --> 00:06:18.716 and check by hand that its leverage matches the balance sheet in that year's 10-K. s12-119 00:06:18.766 --> 00:06:21.632 If that holds, you've got grounds to trust the rest. If it doesn't, s12-120 00:06:21.682 --> 00:06:26.605 you caught a construction error before it went into every table downstream. s12-121 00:06:26.655 --> 00:06:30.850 That check is on you, because only you can decide whether the number looks right. s13-122 00:06:31.600 --> 00:06:35.115 Vague prompts usually don't throw an obvious error. s13-123 00:06:35.165 --> 00:06:39.363 What you get back looks plausible, but it isn't quite what you meant, s13-124 00:06:39.413 --> 00:06:41.840 and that gap grows over a session. s13-125 00:06:41.890 --> 00:06:44.622 So in every prompt you name four things. s13-126 00:06:44.672 --> 00:06:46.326 The file you're operating on. s13-127 00:06:46.376 --> 00:06:49.314 The operation, meaning the transformation you want. s13-128 00:06:49.364 --> 00:06:52.390 The aggregation, the level the result should live at. s13-129 00:06:52.440 --> 00:06:56.681 And the output, by name and format, decided before it runs. s13-130 00:06:56.731 --> 00:06:58.510 The next slide shows what that looks like. s14-131 00:06:59.260 --> 00:07:01.735 Here's the same request written two ways. s14-132 00:07:01.785 --> 00:07:06.483 The vague one: look at this Compustat file and tell me something about leverage. s14-133 00:07:06.533 --> 00:07:08.040 Now the specific one. s14-134 00:07:08.090 --> 00:07:09.398 It names the file. s14-135 00:07:09.448 --> 00:07:13.504 It defines book leverage as long-term debt divided by total assets, s14-136 00:07:13.554 --> 00:07:15.594 for each gvkey and fiscal year, s14-137 00:07:15.644 --> 00:07:18.693 winsorized at the first and ninety-ninth percentiles. s14-138 00:07:18.743 --> 00:07:22.691 It sets the level, the mean by fiscal year and two-digit SIC. s14-139 00:07:22.741 --> 00:07:27.647 And it names both outputs before anything runs, a CSV and a figure. s14-140 00:07:27.697 --> 00:07:31.942 Naming the output is what turns a one-off answer into a script you can rerun, s14-141 00:07:31.992 --> 00:07:34.410 and the Module 2 lab depends on that habit. s15-142 00:07:35.160 --> 00:07:39.318 The failure here is describing the goal instead of the task. s15-143 00:07:39.368 --> 00:07:42.253 Something like, analyze this and tell me what's interesting. s15-144 00:07:42.303 --> 00:07:43.503 It sounds thoughtful, s15-145 00:07:43.553 --> 00:07:47.967 but it hands the agent every decision you should be making yourself: s15-146 00:07:48.017 --> 00:07:51.631 which variables matter, at what level, and how to report them. s15-147 00:07:51.681 --> 00:07:55.408 So before you hit enter, go through the prompt and check for the file, s15-148 00:07:55.458 --> 00:07:58.367 the operation, the aggregation, and the output. s15-149 00:07:58.417 --> 00:08:00.230 If one of them is missing, add it. s16-150 00:08:00.980 --> 00:08:05.591 So let's put this inside an actual research project: idea, data, measure, analysis, s16-151 00:08:05.641 --> 00:08:07.266 paper. s16-152 00:08:07.316 --> 00:08:09.569 The filled boxes are where an agent speeds you up. s16-153 00:08:09.619 --> 00:08:11.107 Brainstorming framings and a quick lit scan. s16-154 00:08:11.157 --> 00:08:14.915 Downloading, cleaning, and joining panels. s16-155 00:08:14.965 --> 00:08:16.291 Drafting extraction pipelines. s16-156 00:08:16.341 --> 00:08:18.458 Running specifications and building figures. s16-157 00:08:18.508 --> 00:08:21.053 Drafting prose and formatting tables. s16-158 00:08:21.103 --> 00:08:25.811 The outlined boxes stay with you: picking the question that matters, s16-159 00:08:25.861 --> 00:08:27.473 judging how the sample is constructed, s16-160 00:08:27.523 --> 00:08:32.212 certifying that a measure captures the construct, s16-161 00:08:32.262 --> 00:08:36.766 interpreting results and choosing the identification, and owning the claims. s16-162 00:08:36.816 --> 00:08:41.330 We come back to this map in Module 6, after you've run the whole pipeline yourself. s17-163 00:08:42.080 --> 00:08:47.784 So, three mantras from this module, and you'll use them for the rest of the course. s17-164 00:08:47.834 --> 00:08:50.231 One, files persist and context doesn't, s17-165 00:08:50.281 --> 00:08:54.590 so anything that has to outlast the session goes in a file. s17-166 00:08:54.640 --> 00:08:56.204 Two, trust but verify, same as an RA. s17-167 00:08:56.254 --> 00:09:01.736 Delegate, but check at least one result yourself before you sign off. s17-168 00:09:01.786 --> 00:09:06.757 Three, be specific: file, operation, aggregation, and output, in every prompt. s17-169 00:09:06.807 --> 00:09:09.738 These go on your verification protocol card, s17-170 00:09:09.788 --> 00:09:13.870 and Module 6 comes back to a harder version of this question. s18-171 00:09:14.620 --> 00:09:17.999 In the live session, the next thirty-five minutes are a demo. s18-172 00:09:18.049 --> 00:09:23.197 I take one task from start to finish in a fresh, empty folder called demo one. s18-173 00:09:23.247 --> 00:09:26.072 I launch Claude Code, go through the permission modes, s18-174 00:09:26.122 --> 00:09:31.206 and type a goal-first prompt that uses the SEC's Financial Statement Data Sets. s18-175 00:09:31.256 --> 00:09:32.946 Then we watch the loop. s18-176 00:09:32.996 --> 00:09:36.621 When it starts going the wrong way, I press Escape and restate what I want, s18-177 00:09:36.671 --> 00:09:39.396 instead of typing corrections on top of it. s18-178 00:09:39.446 --> 00:09:43.585 You want to correct within two turns of noticing the problem, not ten. s18-179 00:09:43.635 --> 00:09:47.179 Then a quick look at Plan Mode, we write a CLAUDE.md, s18-180 00:09:47.229 --> 00:09:50.090 and we check the count against a source Claude didn't touch. s19-181 00:09:50.840 --> 00:09:54.331 When Claude Code wants to run something, like this download script, s19-182 00:09:54.381 --> 00:09:55.689 it can ask you first. s19-183 00:09:55.739 --> 00:10:01.542 You can say yes once, yes and don't ask again for that kind of command this session, s19-184 00:10:01.592 --> 00:10:03.290 or no, and tell it what to do instead. s19-185 00:10:03.340 --> 00:10:04.739 That's Manual mode. s19-186 00:10:04.789 --> 00:10:08.687 It asks before every file edit, command, or network call. s19-187 00:10:08.737 --> 00:10:12.126 Auto hands that review to a second model, a classifier, s19-188 00:10:12.176 --> 00:10:15.521 which lets routine actions through and blocks risky ones. s19-189 00:10:15.571 --> 00:10:18.953 New sessions start in Auto, and Shift Tab changes it. s19-190 00:10:19.003 --> 00:10:21.893 Manual is the safer choice while trust is new. s19-191 00:10:21.943 --> 00:10:25.823 But if you're approving the same kind of action ten or twenty times in a row, s19-192 00:10:25.873 --> 00:10:27.471 you stop reading the prompts. s19-193 00:10:27.521 --> 00:10:31.510 So we use Auto together with git, and git is Module 5. s20-194 00:10:32.260 --> 00:10:36.923 For anything bigger than a single script, ask it to plan before it builds. s20-195 00:10:36.973 --> 00:10:40.402 Shift Tab cycles from Manual to accept edits to Plan Mode, s20-196 00:10:40.452 --> 00:10:43.235 and the status bar tells you which one you're in. s20-197 00:10:43.285 --> 00:10:48.401 In this example the request is a panel of 10-K filing counts by industry and year, s20-198 00:10:48.451 --> 00:10:50.602 from the SEC's quarterly files. s20-199 00:10:50.652 --> 00:10:54.553 Claude reads, checks its assumptions, and proposes numbered steps, s20-200 00:10:54.603 --> 00:10:56.679 with a verification step at the end. s20-201 00:10:56.729 --> 00:10:59.095 Nothing runs until you approve it. s20-202 00:10:59.145 --> 00:11:02.508 Let the plan finish before you decide whether to act on it. s20-203 00:11:02.558 --> 00:11:07.090 Module 3 runs the full cycle of planning, building, and clearing context. s21-204 00:11:07.840 --> 00:11:11.003 Then the lab, eighty minutes in three parts. s21-205 00:11:11.053 --> 00:11:13.912 Part A is twenty-five minutes on your environment, s21-206 00:11:13.962 --> 00:11:19.160 so Claude Code launches from a fresh terminal inside an empty folder called lab one. s21-207 00:11:19.210 --> 00:11:23.746 Part B, fifteen minutes: write your CLAUDE.md, restart the session, s21-208 00:11:23.796 --> 00:11:25.394 and confirm it loaded. s21-209 00:11:25.444 --> 00:11:29.998 Part C, forty minutes: one task on public SEC data, at the base, standard, s21-210 00:11:30.048 --> 00:11:31.040 or stretch tier. s21-211 00:11:31.090 --> 00:11:34.883 Whichever tier you pick, the deliverable is a named script, s21-212 00:11:34.933 --> 00:11:37.427 not an answer in the chat. s21-213 00:11:37.477 --> 00:11:41.070 And you check one output number against a source Claude didn't touch. s22-214 00:11:41.820 --> 00:11:43.523 We close with a debrief. s22-215 00:11:43.573 --> 00:11:44.722 What broke for you? s22-216 00:11:44.772 --> 00:11:47.628 Did you press Escape, or did you argue with it? s22-217 00:11:47.678 --> 00:11:50.280 Whose prompt turned out vaguer than they thought? s22-218 00:11:50.330 --> 00:11:54.367 Then the exit ticket: one thing that clicked, one thing that broke, s22-219 00:11:54.417 --> 00:11:57.084 and one thing you'd want to verify next time. s22-220 00:11:57.134 --> 00:11:58.856 And one piece of homework. s22-221 00:11:58.906 --> 00:12:02.880 Confirm your WRDS account and your Duo enrollment, s22-222 00:12:02.930 --> 00:12:06.186 and test ssh wrds before Module 2. s22-223 00:12:06.236 --> 00:12:07.915 The Module 4 lab depends on that connection, s22-224 00:12:07.965 --> 00:12:11.651 so give yourself at least two days to sort out problems. s22-225 00:12:11.701 --> 00:12:14.147 The night before Module 4 is too late. s22-226 00:12:14.197 --> 00:12:15.210 See you in the lab.