Encoding Domain Knowledge as Agent Infrastructure
Encoding Domain Knowledge as Agent Infrastructure
Source: https://x.com/i/status/2077460395279692197
📌 Teams multiply agent and human output by writing tribal knowledge into the repo as CLAUDE.md/AGENTS.md, skills, and durable automation—so agents work with zero extra prompter context. One-off fixes burn tokens; lint rules, CI, and encoded conventions fix whole classes of problems forever.
🔁 The leverage loop
Classic high-leverage engineering—editor automation, lint rules, CI, e2e tests—matters more with coding agents. Infra speeds humans and fleets of agents; turning a recurring agent fix into a rule or check automates busywork forever instead of re-burning tokens.
📄 Write it down for agents
Agents can navigate unfamiliar codebases, so day-one and non-engineer contributions work only when conventions leave people's heads. Ship CLAUDE.md (Claude Code) and AGENTS.md (cross-tool open format) so agents need zero extra context from the prompter.
🧰 Skills vs always-on instructions
Keep project instruction files lean. Package specialized domain workflows as on-demand Skills (SKILL.md) that load when relevant, preserving context efficiency while still encoding playbooks agents can follow.
⚙️ Rejections are automation failures
“Wrong framework” or “wrong architecture” rejections signal missing automation and docs, not just a bad contributor. Hard gates—types, tests, linters, CI—plus review/docs conventions make the codebase teachable to people and agents.
🌐 Industry traction
AGENTS.md is used across tens of thousands of repos and backed by efforts involving Codex, Cursor, Jules, Amp, Factory, and others. Claude Code pairs human-written CLAUDE.md with auto-memory and relevant skills.
Key facts
| Fact | Value |
|---|---|
| Topic | Encode team domain knowledge as agent-readable infrastructure so coding agents and humans contribute without tribal onboarding |
| Core files | CLAUDE.md (Claude Code session instructions); AGENTS.md (open cross-agent README); Skills/SKILL.md (on-demand domain workflows) |
| Why it multiplies output | Infra speeds humans and agents; durable automation beats one-off token burns; encoded knowledge lowers ramp-up barriers |
| Industry standard | AGENTS.md used by tens of thousands of repos; backed by Codex, Cursor, Jules, Amp, Factory, and others |
| Claude Code memory model | You-written CLAUDE.md rules plus Claude’s auto-memory; skills load when relevant rather than always |
| Source post | Boris Cherny (@bcherny), Jul 15 2026 — every team should ship CLAUDE.md / REVIEW.md / skills / docs for zero-extra-context agent work |
Details
Boris Cherny (Claude Code at Anthropic) argues that high-leverage engineering practices—editor automation, lint rules, CI, and e2e tests—become even more valuable in an agent-driven world. Infrastructure and DevX still accelerate humans, but they also accelerate fleets of coding agents. More importantly, one-off agent fixes waste tokens and miss edge cases; encoding a recurring problem as a lint rule, CI check, or routine permanently removes a whole class of busywork for both people and agents.
What has changed is how much tribal knowledge can live in the repository. Agents can navigate unfamiliar codebases, making day-one and even non-engineer contributions realistic—but only if conventions that used to live in people’s heads are written down as infrastructure. Cherny’s prescription is that teams maintain CLAUDE.md and related docs/skills so agents work productively with zero extra context from the prompter. Rejections for “wrong framework” or “wrong architecture” are failures of automation, not merely of the contributor.
In practice this means lean project instruction files (CLAUDE.md for Claude Code; AGENTS.md as the open cross-tool format), on-demand Skills that package domain workflows, review and docs conventions, and hard automation via types, tests, linters, and CI. Keep always-on instructions short; put specialized playbooks in skills so context stays efficient. The goal matches good senior engineering: multiply output by automating and by making the codebase teachable.