Build Your Own Executive Assistant with Claude Projects, No Code

A Claude project with well-shaped instructions is an internal tool, and you do not write a line of code for it. Each one solves a single specific problem, which is why the right image is not a Swiss Army knife but a Japanese chef's knife: one blade, one job, extremely sharp. I have built around twenty of these, and this walkthrough builds one live, from empty project to working assistant, so the pattern is visible.

The anatomy: instructions plus knowledge

Projects need a paid Claude plan. Create one, name it, and set the instructions; that is the system prompt steering the whole tool. You can get as fancy as you want here, and you should not: start simple, then improve through iterations of testing results. Every good project I have went through many rounds.

The live example is a YouTube packaging assistant. The first instructions were roughly: you are a YouTube expert, you know packaging best practices, write high-CTR titles and meta descriptions, use keywords, keep it short and pleasant to read. Claude already knows a lot about what a good title is from training; instructions steer that knowledge rather than replace it.

Then knowledge files: attach documents to the project, say a collection of titles that made you click, or an expert guide. As I understand it, project files get injected into every chat, unlike the RAG-style file handling in comparable assistant products, where files are chunked, retrieved only sometimes, and you end up begging the tool to please use the knowledge you provided. That retrieval gap is a major weakness of those apps, and the always-present behavior is a real advantage here.

Using and iterating the tool

First run: upload a video transcript, an SRT file straight from the editing tool, and ask for packaging. Out comes a title, description, and summary. Generic, not amazing, and good enough to ship a first version and see what happens, which is the correct bar.

Iteration is just editing the instructions. I added "also give me thumbnail ideas," and the very next run included thumbnail concepts, hooks, and title variations. The thumbnails are a brainstorming crutch; the tool hands you a pile of starting points to refine, and the thinking stays your job.

The second tool: an interviewer that writes scripts

The pattern compounds. My script-writer project has a knowledge file built from a summarized video about scriptwriting, and one instruction that changes its behavior entirely: always collect missing information first. Start a chat and it interviews you, asking the main topic, the target audience, the video type, before writing anything. You can also query the knowledge directly, like asking what video types exist, and it answers from the attached file plus its own training. The result is a complete script from a rough idea and a screenshot.

That collect-inputs-first instruction is the difference between a text generator and something that behaves like an assistant. A tool is only as good as its system prompt, and one sentence of process in the prompt buys more than paragraphs of style guidance.

Why bother when you could code it

I could build all of this as real software. Why would I? The chef's-knife economics only work because a new tool costs twenty minutes: notice a recurring workflow, open a project, write instructions, attach knowledge, iterate. This video is from the Claude Projects era of that habit; the same instinct now feeds skills, which auto-trigger instead of living in a project, and the experiment-first loop for finding tool-worthy workflows is unchanged. Pick the problem you solve every week by hand, and go build the knife for it.