Your senior engineer just rewrote the same Claude prompt for the fourth time this week. Your PM is pasting context into a fresh chat because nobody told her the staff engineer already solved that exact pattern in March. Your designer is one-shotting components with a prompt that took her two months to refine — and it lives in a Notion doc three people can find.
That is knowledge debt. Conservatively, a 12-person product team is burning 8-15 hours a week re-prompting from scratch. At a $150 blended rate, that is $93,600 to $175,500 a year set on fire.
The Org Chart Is Dead. The Context Graph Is Alive.
For twenty years the org chart told you who knew what. People were the index.
When a junior engineer with the right CLAUDE.md, the right agent definitions, and the right retrieved context ships work that used to require a principal engineer, the question is no longer "who do I ask." The question is "whose context am I loading."
The team's expertise is either structured into shared context — versioned, tagged, retrievable — or it is leaking. There is no third option.
The Three Extraction Economies
Their terms, version one: the model ate it. Every prompt sent to a frontier API without a retention strategy is a small donation. The model keeps the aggregate signal. You keep the output for one session.
Their terms, version two: the hourly dump. Mercor and similar platforms pay your best people $80-$200 an hour to dump their IP into someone else's training set. Retail pay for wholesale access to the judgment you hired them for.
Your terms: expert-owned context. The team builds, owns, and versions the library. Prompts, agents, evals, retrieval corpora. Cards stay on your side.
Most teams run one and two by default. Three takes intention.
The Stack
Layer 1: The CLAUDE.md spine. Every repo has a root context file. Architecture, conventions, the do-not-touch list, the test philosophy. Actual, not aspirational.
Layer 2: Role-shaped agents. Reviewer, refactor, migration, copy, research. Each one a markdown file with a defined job, defined inputs, defined output contract. Versioned in git. PR'd like code. When the senior engineer learns a new pattern, she updates the reviewer agent. Everyone is reviewed by her ghost.
Layer 3: The prompt library. Tagged, searchable, indexed. Not a Notion graveyard. A retrieval system.
Layer 4: Eval and feedback. Every failure becomes a test case. The team's mistakes turn into the team's moat. This is the layer most people skip. It is the layer that compounds.
Layer 5: Distribution rules. Who reads what. Who writes what. What never leaves the building. Source code instincts, applied to the upstream artifact that produces source code.
The Math
Team A prompts from scratch. Six turns to usable result. 25 prompts a day. 12 people. 1,800 turns a day rebuilding context.
Team B has the layer. Two turns. 600 turns a day. But Team B is not 3x faster. Team B is compounding. Every output becomes a template. Every failure becomes an eval. Every new hire inherits the library on day one, not month nine.
The Move This Quarter
Audit your prompt surface. Pick one workflow and structure it. Make context contributions a promotion criterion. Decide your terms.
The senior people on your team are senior because of what they know. If what they know is not in the library, they are not actually senior to the system. They are senior to a room they happen to be in.
I wrote a book about this — On Whose Terms: The New Expert Economy and the Fight for What You Know. Grab a copy.
The org chart is being rewritten in markdown. Either you are writing it, or your tools are writing it for you.
Keep Building,
— Matt
Matt Cretzman builds the Knowledge Delivery System at Skill Refinery and writes about the expert economy at mattcretzman.com. Author of On Whose Terms.
Top comments (0)