Every knowledge worker has experienced this: you solve a complex problem, navigate a tricky edge case, or figure out the perfect workflow for a recurring task. A month later, you face the same situation and realize you’ve forgotten half of what you learned. The expertise you built evaporated into the ether of your memory.
This happens because the most valuable knowledge we develop isn’t easily captured in traditional documentation. It’s not the what—the steps we took. It’s the why behind each decision and the how of our thinking when things don’t go according to plan.
The Missing Layer in Enterprise Systems
Jaya Gupta recently published a piece on Foundation Capital called “Context Graphs: AI’s Trillion-Dollar Opportunity” that articulates this problem beautifully at the enterprise level. Her core argument: companies have invested billions in systems of record—Salesforce, Workday, SAP—but these systems capture data, not reasoning.
The actual operating logic of an organization? That lives in Slack threads. In escalation calls. In the heads of senior employees who “just know” how to handle edge cases. When an AI agent encounters an exception, it can’t tap into this tribal knowledge because it simply doesn’t exist in any structured, accessible form.
Gupta’s focus is enterprise infrastructure, but I’ve found the same principle applies at the individual level. We all have our own “tribal knowledge”—hard-won expertise that we carry in our heads but never externalize in a way that’s actually useful for future reference.
The Problem with Traditional Documentation
Most of us have attempted to solve this with documentation. We write SOPs, create runbooks, build wikis. But here’s the uncomfortable truth: these artifacts often capture the actions without capturing the intelligence.
A standard operating procedure might tell you: “Step 1: Open the CRM. Step 2: Filter by X criteria. Step 3: Export to CSV.” What it doesn’t tell you is:
- Why you’re filtering by that specific criteria instead of alternatives
- What to do when the data looks “off” in ways that are hard to articulate
- How to recognize when this process is the wrong one for the situation
- What “good” looks like at each stage so you can course-correct
This is the tacit knowledge problem. The stuff that makes experts experts isn’t their ability to follow steps—it’s their judgment, pattern recognition, and decision-making under ambiguity. And that’s exactly what’s hardest to document.
Building Skills, Not Just Procedures
I’ve been experimenting with a different approach using Claude Code. Rather than trying to document my work after the fact, I have Claude extensively log my actions, my rationale for decisions, and the outcomes in real-time as I work. This creates a rich record of not just what I did, but how I was thinking.
From these logs, I create what Claude calls “Skills.” But Skills aren’t just SOPs for AI agents. They’re something fundamentally different—they capture how I would think about a problem:
- Decision logic: Not just “do X,” but “when you see A and B but not C, consider doing X because…”
- Edge cases: The exceptions I’ve encountered and how I’ve handled them
- Tool orchestration: How different tools and systems work together, including their quirks and limitations
- Quality criteria: What “done well” actually looks like, in specific and observable terms
- Failure modes: What can go wrong and how to recognize when it’s happening
This is tacit knowledge, captured from actual task execution rather than reconstructed from memory.
From Christmas Cards to Sales Operations
The range of Skills I’ve built might seem eclectic, but that’s precisely the point. Real expertise is domain-specific and often surprisingly narrow in its applicability.
I have Skills for operationalizing outbound sales plays using the Salesforce CLI and Salesloft API—knowledge about which fields matter, how to structure sequences, and what conversion patterns I’ve seen work. I have Skills for building consistent UIs across all my products, capturing not just design system rules but the reasoning behind when to break them. I have Skills for setting up both basic and complex reporting, including the edge cases that tend to trip people up.
And last week? I built a Skill after working with Claude Code to properly format Christmas card mailing labels. It’s a task I do once a year, and I had wasted 30 minutes last December trying to remember the specific printer settings, margin adjustments, and data formatting that actually work. Now that expertise is captured, ready for next December.
The Compound Effect
Here’s what changes when you start building your personal context graph:
A new outbound sales play now spins up in minutes instead of hours. Not because I’m following a checklist, but because the nuanced understanding of how to do it well—the judgment calls, the quality checks, the “watch out for this” warnings—is all accessible and actionable.
The expertise that Claude and I built together during previous work isn’t lost. It’s codified in a format optimized for future AI-assisted task execution. Each time I face a similar challenge, I’m not starting from scratch or relying on my imperfect memory. I’m building on a foundation of captured intelligence.
The result is working faster, with higher quality, and with consistent results. Not because the work is “automated” in the traditional sense, but because the hardest part of complex work—the thinking—is preserved and reusable.
The Broader Implication
Gupta is right that enterprises need infrastructure for capturing context at scale. But individuals don’t have to wait for that infrastructure to exist. You can start building your own context graph today, one Skill at a time.
The key insight is that the best time to capture tacit knowledge is while you’re actively using it. Document your thinking in real-time, not after the fact. Let AI help you externalize the reasoning that would otherwise remain locked in your head. Then structure that knowledge in a way that’s actually useful—not as passive documentation to be read, but as active Skills that can inform future work.
The future of AI assistance isn’t about agents following rigid instructions. It’s about agents that understand not just what to do, but how you think about doing it. Building that understanding starts with capturing the context that currently exists nowhere except in your own expertise.
The referenced article, “Context Graphs: AI’s Trillion-Dollar Opportunity” by Jaya Gupta and Ashu Garg at Foundation Capital, explores the enterprise infrastructure implications of this problem in greater depth.