Sorry for the delay, y’all. I’m now posting the documents that I used to build my Tiller “agent”.
You can download the three documents from here: Notion – The all-in-one workspace for your notes, tasks, wikis, and databases.
You should know up front that I’m using Anthropic’s Claude with the “3.5 Sonnet” model and that I have their $20 per month “Professional” subscription which enables their “project” feature. A project lets you upload document files to serve as “knowledge” for prompts and these stay around across sessions.
You may be able to use their free version to get results by pasting my documents into the prompt itself, but that’s iffy because together they’re pretty large and may exceed the “context window” size for a prompt.
My advice: Sign up for a month of of Claude’s professional plan to play with it and cancel before it renews. I use it constantly for work and personal things and I think it pays for itself quickly.
Documents
TillerProductSummary.txt
This is a page of text that summarizes what Tiller is. I think I just selected and copied the text from a page on the Tiller website and pasted it into this text file. Not sure how important this is, but you might as well include it.
CurrentSheets.txt
This is the big one. It has all the information I could find on each Tiller sheet I’m using, copied from the website and just pasted into a big, long, ugly-ass text file. This is my “knowledge” document. Each time I prompt my agent, it’s using this as its “smarts” together with whatever other knowledge it already has about Tiller from scraping the internet. I did give the document a little structure, breaking it into sections for each of the Tiller sheets that I’ve started with. As I tell the AI, these sections are bounded with special character sequences.
Prompt.txt
This is the prompt that I started my conversation with. This is where some of the magic is. I’m using some important prompting techniques here. I’m defining my goals, my background, the level of my expertise, and explaining what I’ve done so far. Then I tell the AI who it is, how I want to work with it, and I ask it to read my documents. An important thing: AI’s are trained to be subservient, overly-nice, praiseful, and to follow your lead way too much. You need to tell it to not be nice and push back if it thinks it’s right. You’ll want to customize this file for your own self and goals.
Steps
- Log into your Claude account and go to Claude
- Click on the “Claude” text at the upper left of the window
- Click the “Projects” item to reveal the “Projects” page
- Click the “Create Project” button at the upper right of the window.
- Enter your project name.
- Enter a short description. Something like “Getting my finances in order using Tiller spreadsheets”.
- On the new project page, click the “Add Content” button and upload the “TillerProductSummary” and “CurrentSheets.txt” files.
- Make sure you’re using the “3.5 Sonnet” model. Look for the popup button at the bottom left of the white prompt field.
- In the prompt field (labeled “How can Claude help you today”), paste in the contents of the “Prompt.txt” file and click the up-arrow button to submit it.
At this point you’re ready to start a conversation. Having some previous experience with prompting will be helpful. Basic advice: Pretend you’re talking to a real human being —an idiot savant who is exceptionally brilliant and insightful in some ways and utterly clueless in others. The quality of the help you get will depend on the quality of your prompts. Provide as much context as you can in your questions. Be verbose. You need to be on guard for hallucinations. Challenge things that don’t seem right. Ask to break things down into steps. Take screenshots of your sheets and add them to your prompts.
Happy to answer questions here
Aside: I’m a software engineer with some experience building AI pipelines. If anyone is interested in building a proper agent, let me know.