Curate — The Context Your AI Agents Consume

The first opens the tool and starts prompting. The AI doesn't know your naming conventions, your service-layer rules, which database schema owns what, or the three patterns your team has agreed never to use. So it guesses. The engineer corrects it. It guesses again. By end of day there's a pull request that compiles, violates two architectural rules, and will get sent back in review.

The second engineer opens the same tool. The AI already knows the conventions, the patterns, the schema ownership, the "never do this" list — because it's all in a file the AI loads automatically at the start of every session. The first prompt lands close to correct. The engineer refines, not re-explains. Ships by lunch.

Same model. Same task. Same licence fee. The gap between them is not talent. It is context discipline — and that is a property of the organisation, not the individual.

Anthropic calls it context engineering: "curating and maintaining the optimal set of tokens during LLM inference." Curate is the org-wide version of that discipline — you treat the context your AI consumes as a first-class engineering asset, versioned and reviewed like code, not something each developer reassembles from memory every session.

AI-readable standards. The AGENTS.md file has become the universal, cross-vendor way to tell an AI agent how your codebase works — adopted across 60,000+ open-source repos and supported by Claude Code, Cursor, Copilot, Codex, Gemini and more. CLAUDE.md adds Anthropic-specific power. These are instructions written for the agent, checked into the repo, reviewed in the PR.

Layered context with nearest-wins precedence. Org-level rules, then repo, then service, then feature — the most specific file wins. A five-engineer team runs this at low resolution; a five-thousand-engineer org runs the same shape at high resolution. The structure doesn't change with size. The resolution does.

Indexing. Automated, semantic indexing of the codebase — dependency graphs, cross-service awareness — so the AI can find the right code instead of starting cold on every prompt.

The counter-intuitive part: context quality beats context volume. Stanford's "lost-in-the-middle" research shows model correctness drops off once you push past long contexts — dumping your whole wiki into the prompt makes the output worse, not better. Curate is about feeding the right tokens, not the most.

The gap between your best and worst AI users is not a hiring problem. It is a context problem, and you already own it.

The evidence is measurable. Time-to-10th-PR for new hires dropped from 91 days without AI to 49 days with daily AI use — a 46% reduction (DX 2025) — but that number belongs to orgs with the context discipline to make AI useful from day one. Without it, roughly 72% of organisations need a month or more just to get developers productively onboarded to AI (industry surveys). And Stanford's lost-in-the-middle finding means the naive fix — feed the AI everything — actively degrades output.

So the org that "adopted AI" but never curated context is paying full price for the tools and collecting a fraction of the value — while believing the shortfall is about which engineers it hired.

The ladder runs: The Ad-Hoc Org → The Individual Champions → The Emerging Framework → The Structured Org → The AI-Native Org.

The Ad-Hoc Org starts every prompt cold. The Individual Champions has a few developers who've cracked it privately — and loses that knowledge the moment they leave. The Emerging Framework has some AGENTS.md files but no governance, so adoption is uneven. The Structured Org has standards documented and followed across teams, with output that's finally consistent. The AI-Native Org treats Curate as a first-class capability — measured, maintained, and retained when people move on.

The jump that matters is from "a few people figured it out" to "the org figured it out." That is what turns a 10x team from a personality into a practice.