Context & Configuration — AI Base Knowledge

The configuration stack

Five layers, one mental model.

The clearest way to know which mechanism to reach for. Each layer answers a different question.

1 · Custom instructions

The always-on baseline — for example, use TypeScript and prefer functional components.

2 · AGENTS.md

Project context — the minimum viable onboarding for your repo.

3 · SKILL.md

On-demand expertise, loaded only when relevant.

4 · MCP servers

External tool and data access.

5 · Cursor rules

Editor-specific behavior.

AI Skills

Custom onboarding that loads on demand.

Skills are folders with a SKILL.md file (name + description in YAML) plus optional scripts. They use progressive disclosure: only the name and description load at startup; the full body loads only when the model judges the skill relevant — saving context.

📚

Encode expertise

Code review, deploy checklists, DB conventions — repeatable team knowledge without re-prompting every session.

🌐

An open standard

Published by Anthropic on Dec 18, 2025 (agentskills.io) — portable across Claude apps, Claude Code, and the API.

🛡️

Install from trusted sources

Anthropic warns a malicious skill can direct an agent to exfiltrate data. Note: Skills are not ZDR-eligible.

Cursor Rules

Pin conventions in the editor.

Cursor reads persistent rules from .cursor/rules/*.mdc, with frontmatter for glob-scoped, apply-intelligently, or manual activation. Precedence is Team → Project → User. Rules pin framework versions and guardrails so the agent stops repeating mistakes.

CLAUDE.md & AGENTS.md

One markdown file onboards every agent.

CLAUDE.md is read at the start of every Claude Code session. AGENTS.md (OpenAI, Aug 2025) became the cross-tool standard — adopted by 60,000+ open-source projects by Dec 2025 and contributed to the Linux Foundation. Keep it short (~200 lines), imperative, and version-controlled.

Strong, direct wording like NEVER and YOU MUST measurably increases how well an agent follows your rules.

Keep project files short, imperative, and in version control — the nearest nested file wins.

Custom instructions & system prompts

Shared with the team, or set per user?

The distinction governs what's visible and versioned versus what's private.

	System prompts	Rules files
Sets	Persona & constraints	Repo conventions & guardrails
Visibility	Usually invisible	Visible repo artifacts
Shared with team	Often per-user	Version-controlled, shared