Building a Predictable AI Office for Consistent Delegation
Source: Social Media Examiner
BuildingPredictableOffice
Tired of re-explaining your processes, templates, and preferences every time you open ChatGPT or Claude? Wondering how to set up an AI workspace that lets you delegate real work to AI and get the same reliable result every time? In this article, you’ll discover how to build an AI workspace, or AI office, of process documents, templates, and historical context that ChatGPT, Claude, and other AI tools can draw on to complete delegated work consistently. This article was co-created by Tiffany Kyazze and Michael Stelzner. For more about Tiffany, scroll to the end of this article.
Why an AI Workspace Makes Delegation Possible
Tiffany Kyazze came to AI through automation, which she defines as sequencing: when this thing happens, then that thing happens. In her AI Flow Club community, she teaches entrepreneurs to build deliberately boring AI systems that are well-thought-out and vetted once, then deliver results they can count on. Because AI doesn’t always produce a deterministic result, her goal is to make it controlled, contained, and profitable. Work can only be handed off to AI when the output is predictable. Tiffany compares the ideal to a McDouble from McDonald’s: it’s the same every time, and that’s the point. An AI workspace provides that predictability by giving AI a single place to find the standards, processes, and templates for every task it’s asked to do.
#1: Give AI Only the Context Each Task Needs, Starting Small
The most common misconception Tiffany sees is the belief that AI needs to know everything: personal motivations, how someone works, what’s happening in their life, and data from every app they’ve ever used. The connectors and plugins pages in ChatGPT and Claude make this tempting because so many familiar apps are listed there. The problem is that a large language model assumes anything brought up in a conversation is meant to be used. If someone tells AI they’re feeling tired and asks it to write an email declining a podcast appearance, the email will likely say the sender is tired, when the sender may have wanted it to say only that the timing doesn’t work. Connected data carries the same risk. A Stripe connector can help reconcile revenue, but if AI knows a business makes $10,000 a week, that figure could find its way into social media posts. Second brain systems are useful for knowing a personal story, but according to Tiffany, AI doesn’t need that story to do work. For each task, the right questions are which information is relevant and what the risk is if AI overshares it. Precise context also saves tokens. Without a template, a request to summarize a meeting forces AI to decide what a summary looks like, what to include, and what counts as valuable. The result is a longer document filled with information that’s relevant but not needed. If the summary isn’t told to focus on project advancement, for example, the five minutes participants spent talking about their weekend will end up in it. Tiffany recommends defining both the type of information needed and the structure of the output. That definition also prevents AI from reading more than it needs: to summarize a call, it uses only the call transcript and the call summary template, not unrelated files such as scheduling rules.
How to Start an AI Workspace Without Overloading It
A popular trend has been to use the master prompt, an autobiography provided to AI with every request. Some people wrote 90-page master prompts. Tiffany says this was too much information, because AI doesn’t pay close enough attention to small details buried in a huge document. It’s also hard to maintain: if a brand voice needs to change, finding that section and every other section influencing it takes a lot of time. Instead, Tiffany recommends starting with just a couple of pieces and making sure each one works properly. Starting small makes it possible to tweak each addition and see exactly how AI uses that information. With 55 documents in the folder on day one, it’s hard to tell which ones AI referred to when generating a result. Anyone who already has that many should explicitly tell the AI which documents to use for each task. Then, if the result is poor, there are only two places to check: the instruction or the context. A Claude project that already performs well is a good first candidate. A project is a folder with instructions and files focused on one task, so it can become a subfolder of the new workspace, where any AI tool can access it.
#2: Choose Where the AI Workspace Lives
An AI workspace is a folder on a computer with information organized into subfolders. Tiffany describes it as a digital binder, like the binder she received at her first job, with tabs explaining how to handle each situation, such as a call to the front desk. Instead of tabs for administrative information, processes, and templates, the workspace has subfolders, and AI knows exactly where to find each one the next time a task comes up. The workspace works with tools that can access files, such as Claude Cowork, Claude Code, ChatGPT Work, and ChatGPT Codex. According to Tiffany, regular ChatGPT also has access to the file system. Google Drive can serve as the source of truth, too, since it connects to nearly every AI tool.
How to Set Up the Workspace With Markdown and Obsidian
Tiffany keeps her workspace in her computer’s file system as Markdown files. Each file is only a couple of kilobytes, which keeps token use low. She calls Markdown the first language of AI. Even unformatted, it’s easy to read: a hashtag marks a heading and asterisks mark bold text. For anyone who doesn’t want to read raw Markdown, Tiffany recommends Obsidian, which displays the files with clean formatting and shows which documents are connected. Those connections let AI pull everything a task needs. An email process, for example, can link to example emails, an email template, and the rules for deciding which messages deserve a response. Tiffany prefers Obsidian to Notion because AI locates information in it faster. A Notion search in Claude scans many fields, and her Notion databases sometimes have 20 or 30 columns. Obsidian stores the same kind of fields in YAML, a small block of code at the top of each file, and still displays the file like a database.
How to Sync the Workspace With a Team Using GitHub
Tiffany syncs her workspace with GitHub. She keeps a folder for her team office on her computer, and her partner Taylor, who lives in Canada, keeps a matching folder on hers. When Tiffany pushes changes to GitHub, Taylor’s folder updates automatically, and if both edit the same document, GitHub merges the changes without one overwriting the other. GitHub also gives new AI tools access without handing over the file system. When Muse launched, Tiffany pointed it to her GitHub repository instead of giving it access to her computer, and it got the information it needed.
#3: Fill the AI Office With Process Documents, Templates, and Historical Context
Tiffany calls the finished workspace an AI office: a folder of files that AI uses as a starting point. It holds three types of content.
- Process Documents: A process document outlines the steps to take and serves as a standard operating procedure. Tiffany’s standards for deciding which emails are worth responding to are one example.
- Templates: A template is either an example of a finished output or a document that AI duplicates and fills in. A fill-in template might include only headings with