AI Fundamentals

Here's what a bad prompt looks like. Can you write me a draft of my quarterly LP letter? Claude tries its best, but with no context, it makes a generic output that the user still has to spend hours customizing. The problem isn't Claude, it's the prompt.

I'm Alex, and I run our change agents program here at Carta. That means I spend a lot of time on-site with our venture capital and private equity customers, helping them embed AI into their day to day work. Today, we're covering two things, how to write prompts that actually work and how to turn them into repeatable workflows your whole team can run. Here's what we'll cover.

The anatomy of a prompt that gets results, system prompts, the standing instructions that shape every session, and how to build skills and artifacts that your whole team can use. Ready? Let's get into it. Prompting.

Every good prompt has four parts. Context, goal, format, and constraints. Context is what Claude needs to know about the situation. Goal is what you actually want.

Format is how you want the output, and constraints are what Claude should never do. Here's a weak prompt, summarize our fund performance. But we can turn it into a strong prompt: the Carta Investors plugin, pull NAV, IRR, and TVPI for fund three as of Q1 twenty twenty six. Summarize performance in three bullet points for my LP meeting with Emily Wilson tomorrow.

Cite the source and ask me any clarifying questions as needed. Same request with much more predictable output. Now, one pattern that can give you even stronger results. When you're doing something recurring, a quarterly report, an LP update, try this.

Give Claude the prior output, like an Excel report, the inputs that produced it, like source documents, and the new inputs. Then just ask it to reproduce that same output for a new period, a new LP, or a new font. This forces Claude to use the new inputs rather than pattern matching from the old output. It's the most reliable prompting pattern for recurring financial work, and it can be as easy as taking a screenshot to show Claude a previous output.

Still with me? Good. Now let's talk about how to reduce time giving Claude context and instructions. System prompts, making Claude yours.

A system prompt is a set of standing instructions that runs before every session. It's like onboarding a new analyst. These are the pieces of feedback you find yourself giving over and over again. Here's a few simple examples we recommend when we're on-site with customers.

Be concise. Fewer, clearer sentences. Cite the source and as of date on every financial figure. Ask clarifying questions before starting any complex task.

That last instruction about clarifying questions is especially important. It prevents Claude from going off and building the wrong thing for twenty minutes before realizing it misunderstood the goal, which saves you time and tokens. You can go to your user settings to enter your system prompt. We've linked a starter system prompt at the end of this video.

You can paste it directly into your cloud settings. After your starter prompt, we suggest you build your system prompt iteratively. After any complex task, ask Claude to suggest updates to your standing instructions based on what it just meant. Building repeatable workflows.

Now let's turn great prompts into something the whole team can run. There are two primary ways to make your work in Claude more repeatable. First is a skill or a saved set of instructions. These are the recipes that we mentioned in video two.

Claude writes it once and uses it every time you need to complete that same task. The output is extremely flexible. It can be a Slack message, a spreadsheet, or even a branded PDF. Some skills that we've built in recent client Onsites include compare a Portco's board deck to previous quarters to find KPI changes, download invoices from your email, and allocate expenses between funds and your manco, and LP meeting prep tear sheets, including performance summary, CRM notes, and recent valuation changes from Cartag.

How to save a workflow as a skill. Here's the simplest path. After a session that produced the output you wanted, go back to that Claude window and type package this up as a skill, then hit the save skill button. It appears in settings under skills and you can then share it with anyone on your team.

They can install it and run the exact same workflow. Artifacts. An artifact is an interactive view like a live dashboard or chart that refreshes when your data You can return to it over and over again. And while creating one cost tokens, refreshing it does not.

Artifacts are the right pattern to use for data you're consistently being asked for by LPs or by your GP. Here's an example that comes out of the box from the plugin. This will look very similar to the SOI in the Carta web app, but in addition to quick access, there's big advantages here. Artifacts are almost endlessly customizable.

If you want to add a drop down to select a different period, add a column to see valuation changes, or even make the formatting fit your firm branding, you're just a Claude Prompt away. Artifacts are also shareable, meaning you can share Carta data internally without giving access to the whole Carta web app. Share the link once with your colleague and a refreshable artifact shows up right in their Cloud instance. A quick recap.

Every prompt needs context, goal, format, and constraints. Give Claude a previous example when automating an existing process, and set a system prompt once that will run every session. For repeatable work, build skills for recurring tasks that produce documents and artifacts for data you want to explore visually and revisit. In the next video, we teach you how to use the toolkit you've learned so far and zoom out from individual tasks to rethinking how your firm operates from the ground up.

When you're ready, click below.