The Gap No Prompt Can Close
Why the middle layers get skipped, and how to name what Claude can't know about your work.
TUTOR WITH THEFOCUS.AI
Copy this prompt into Claude, ChatGPT, or any external AI assistant. It points the assistant to the course instructions and links it to your student profile to track your progress and customize observations.
You are not enrolled yet. Enroll to generate a Student ID to track lesson completions and store learning notes.
The middle layers get skipped
A product leader watches an agent demo. Autonomous coding. Multi-step research. A system that deploys its own work, monitors the results, and adjusts. She looks at it and then looks inward: her team’s processes run on habit, knowledge lives in three people’s heads, and the last time someone tried to document a workflow it was a Google Doc that nobody updated.
This is the gap you just felt in the last lesson’s exercise. Not a technology gap — an organizational one. Most companies want to skip from scattered ChatGPT use directly to autonomous agents. The middle layers get skipped. But the middle layers are where all the value lives.
You cannot delegate to a system that does not understand your work. You cannot trust a system that cannot show you where its answers came from. You cannot let a system act autonomously if nobody has defined the boundaries. And you cannot close the feedback loop if there is no infrastructure to capture what worked and what did not.
These are organizational problems, not AI problems. The technology is ready. The organizations are not.
We spent six weeks at a bookkeeping firm before writing a single line of agent code. Six weeks just documenting business logic that had never been written down. The bookkeepers absorbed ambiguity that no system could replicate without explicit rules — which customer uses which chart of accounts, which exceptions get escalated, which transactions get categorized differently in Q4. They had been doing this by feel for years. It worked because humans are remarkably good at absorbing context without formalizing it. But you cannot hand that to a machine.
At a construction company, a single developer made 224 commits over three months encoding cost codes that lived entirely in one person’s head. If that person left, the company would have lost the mapping between their accounting system and reality. Two hundred and twenty-four commits. Not building an AI system — just writing down what one human already knew.
Sarah at Cornwall Market has her own version of this. The categorization rules for fifteen suppliers live in her head. When an invoice comes in from Chen’s Produce, she knows without thinking that everything goes to account 5100. When one comes from the regional distributor, she mentally splits items across four different cost-of-goods sub-accounts based on department. When there is an oddly large charge from a supplier, she knows which ones round up (Chen’s always does, it is fine) and which ones warrant a phone call. None of this is written down anywhere. If Sarah got sick for a month, her staff would muddle through, making mistakes she would spend weeks cleaning up.
That is the gap. And it is universal. Every organization has it.
Name what is missing
Go back to the exercises from the first two lessons. You asked Claude a real work question, then asked it again with proper context. Some gaps closed when you attached documents. Some did not. The ones that did not close are your legibility gap.
Write down three of them. Be specific. Not “it didn’t understand our business” — that is too vague to act on. What exactly did it not know?
Sarah’s legibility gap list for Cornwall Market looks like this:
-
Custom chart of accounts mapping. Cornwall Market uses sub-accounts that their accountant set up specifically for their business. Account 5100 is Cost of Goods — Produce. Account 5120 is Cost of Goods — Bakery Ingredients. Account 5200 is Cost of Goods — Beverages. Claude knows what a chart of accounts is. Claude does not know this chart of accounts — unless Sarah attaches it to every single conversation.
-
Q4 flagging rules. Starting in October, any contractor payment needs to be cross-referenced against year-to-date totals for 1099 reporting. Sarah does this mentally — she just knows which contractors are approaching the threshold. Claude has no idea what Cornwall Market has paid any contractor this year.
-
Supplier name ambiguities. “Chen’s” could be Chen’s Produce (account 5100) or Chen’s Bakery Supplies (account 5120). Sarah knows which is which from context — the invoice format, the items listed, the delivery day. A new hire would not. Claude definitely would not.
These three items share a pattern. They are not secret. They are not complex. They are just undocumented. A new hire at Cornwall Market would learn them eventually through experience, asking questions, and making mistakes that Sarah would catch. That learning process takes weeks or months. And it is exactly the same learning process that an AI system needs to go through — except that an AI system cannot learn by osmosis. It needs the knowledge to be explicit.
Write down three things Claude got wrong (or could not possibly know) even after you gave it context. For each one, ask yourself: would a new hire know this on day one? If not, how would they learn it? The answer to "how would they learn it" is exactly the knowledge you need to capture.
Keep this list. You will start capturing it by voice in Chapter 02, and turn it into a skill in Chapter 04.
Here is what you have done in this chapter: you have learned your way around Claude, felt the difference context makes, and felt the gap that context alone cannot close. You have located yourself on the maturity framework. You have a map of the four transitions you will walk through. And you have a concrete list of knowledge gaps to work with.
That list is the starting point for everything that follows. In Chapter 02, you will get Claude onto your phone and learn the fastest way to capture what you know: by talking.
Check your understanding
Answer in your own words — write it down before opening the key. Your tutor grades against the criteria and generates fresh variants on retries.
Q1. You attached your chart of accounts and got a much better answer — but attaching documents to every conversation cannot close the whole gap. Name the two kinds of organizational knowledge that stay out of reach of the attach button, with an example of each.
Q2. “Most companies want to skip from scattered chat use straight to autonomous agents.” Explain why the middle layers cannot be skipped — what specifically breaks when an organization tries?
Answer key — attempt every question first
Answer key
Q1
Model answer: First, system-resident data — running totals, live records, history that sits in QuickBooks or a CRM and changes daily; no attachment stays current. Second, tacit rules — the exceptions and judgment calls (“Chen’s always rounds up, it’s fine”) that nobody has ever written down, so there is nothing to attach.
Pass criteria: identifies live/system data as one category; identifies undocumented/tacit rules as the other; gives a plausible example of each
Q2
Model answer: Each layer is load-bearing for the next. You cannot delegate to a system that does not understand your work (no legibility), cannot act on answers you cannot verify (no trust), and cannot grant autonomy without defined boundaries. Skipping to agents on top of tribal knowledge produces confidently wrong automation — generic answers executed at machine speed.
Pass criteria: articulates the dependency chain (legibility before trust before delegation); states a concrete failure mode of skipping, not just ‘it fails’