Knowledge Priming: The Attention Budget

Knowledge as code fails in production when the files exist but the agent still ships the average of the internet.

In Knowledge as Code, I argued for the file: institutional memory that survives a tool change because it is reviewable text, not a trapped binary. In Negative Knowledge, I argued for the retirement trail: the audit of what you already rejected. This essay is about the attention budget. What does the agent read first, and who decided that?

The Demo, Then the Market Signal

A few weeks ago I walked operational teams through a knowledge system I build with markdown in source control. They were experimenting with agentic workflows. The part that landed was not a model choice. It was the substrate: project context, rules, and standards living as portable files instead of opaque packages nobody can review.

Four days later, Google’s Open Knowledge Format (OKF) made a similar bet public: vendor-neutral markdown plus front matter for knowledge-intensive work. The timing mattered less as coincidence than as confirmation. Teams were already hungry for a contract that agents and humans can share.

Building those workflows has been a cycle of learning and relearning, especially now that token budgets are a boardroom phrase and not only an engineering one. The scar tissue looks like this: you can hand an agent every file you own and still watch it produce internet-average prose, because nothing governed what loaded first.

Priming Is an Attention Budget

Rahul Garg names the coding-side practice Knowledge Priming in an essay published on Martin Fowler’s site: treat project context as infrastructure, a curated onboarding packet that overrides training-data defaults before the first prompt. His target shape is lean, often under fifty lines of essential context, with anti-patterns listed as carefully as the stack versions.

I extend the same System-layer move beyond code generation. Thought leadership, proof systems, and genealogy research all fail the same way when the corpus is present and the primer is not: the model fills the gap with the average of the internet. Voice drifts. Claims soften. Recovery costs more than a thin file would have.

Priming is not more documentation. It is curation with a budget. Within Knowledge as Code, OKF-shaped knowledge bundles are a strong vehicle for that budget: a self-contained, agent-parseable unit that carries boundaries, vocabulary, schemas, and team standards into the start of a session. The bundle is the mental model you would hand a new hire on day one, written so a machine can digest it without wading through historical noise.

Three design moves keep the budget honest:

1. Lean always-on context. Put the orientation in a thin primer. Keep long procedures in on-demand skills. A six-hundred-line instruction file is not thoroughness; it is context flooding dressed as governance.
2. Recoverable state. Capture enough session and project state that continuation does not require re-explaining the world each morning.
3. System and team guardrails. Inject conventions and standards as first-class primer content, so alignment is not a hope attached to the last chat message.

Together those moves shrink Cognitive Debt on both sides of the review loop: the human stops re-teaching the house style; the agent stops inventing a generic house style of its own.

Own the Onboarding Packet

If every new hire gets a manager and an onboarding packet, your agent deserves at least the packet. The open question in most organizations is ownership. Who curates the always-on layer? Who retires stale rules? Who decides what stays out of the window so the verified source of truth can actually steer attention?

Write the fifty-line primer before you buy another platform. Bundle the rules with the background. Point the agent at trusted sources, then tell it what not to do. Treat that file like infrastructure: versioned, reviewable, boring on purpose.

The files were never the finish line. The attention budget is.

Who owns the agent’s onboarding packet in your organization, and why is the role still unassigned?

Madam I’m Adam

This continues the thread from Knowledge as Code and Negative Knowledge: the file and the retirement trail only hold if you also govern what the agent reads first.