What is context engineering for Claude?

Context engineering is the discipline of systematically managing everything that enters Claude’s context window — instructions, domain knowledge, tools and memory — rather than just crafting individual prompts. Anthropic describes it as the natural progression of prompt engineering: prompt engineering is about how you ask, context engineering is about how you structure the workspace (AGENTS.md, Projects) and what you connect (MCP servers, Skills, Memory). In the enterprise it becomes a software-engineering practice with version control and auditability.

How is context engineering different from prompt engineering?

Prompt engineering optimizes a single request — the wording, examples and structure of one instruction. Context engineering optimizes the whole environment the agent operates in across many turns and sessions: the repository manifest (AGENTS.md), reusable Skills, connected MCP tools, persistent Projects and durable Memory. Static prompts do not scale to agentic workflows, so teams move from "writing the perfect prompt" to "architecting the context" — assembling the right information dynamically from multiple sources.

What are the core building blocks of enterprise context engineering with Claude?

Eight building blocks: (1) AGENTS.md / CLAUDE.md — the repo-level system manifest; (2) Agent Skills — on-demand domain expertise; (3) MCP servers — the data and tool integration layer; (4) Claude Projects — shared workspaces with a knowledge base; (5) the Memory tool — durable cross-session context; (6) Plugins — packaged, distributable bundles; (7) slash commands and routines — reusable workflows; (8) Subagents — context isolation and parallelism. Most enterprise rollouts combine AGENTS.md + Skills + MCP first, then add Projects, Memory and Plugins as they scale.

What is AGENTS.md and why does it matter for teams?

AGENTS.md (and Claude’s CLAUDE.md) is a markdown manifest at the root of a repository that acts as a persistent, project-level system prompt — defining the agent’s identity, coding conventions, behavioral constraints and workflows. Because it lives in git, it is reviewed, versioned and consistent for every developer and every session. It is an open, cross-tool convention (supported by Claude Code, GitHub Copilot, Cursor and others), which is why it is the foundational, lowest-cost context-engineering step.

What is the role of MCP servers in context engineering?

The Model Context Protocol (MCP) is the open standard — now governed by the Linux Foundation — that connects Claude to enterprise data silos and tools such as Jira, GitHub, SQL databases and internal APIs. Instead of pasting data into prompts or maintaining brittle bespoke integrations, MCP servers expose typed tools over stdio or Streamable HTTP, giving agents a reusable, least-privilege access layer. MCP is the "last mile" that turns Claude from a chatbot into a system that can act on your real data.

How do Skills, Plugins and Subagents fit together?

Skills are reusable packages of domain expertise that load on demand. Plugins are the distribution mechanism: a plugin bundles Skills, slash commands, MCP servers and hooks so an enterprise can ship a curated context setup to every developer in one step via a marketplace. Subagents are the orchestration primitive: isolated agents with their own scoped tools and fresh context that handle narrow tasks and protect the main context window from overflow. Together they let you package, distribute and scale context engineering across a team.

How should an enterprise start rolling out context engineering for Claude?

Begin with an AGENTS.md manifest in your key repositories — conventions, guardrails and identity — and version it. Add Skills for repeatable expertise and connect 2-3 high-value MCP servers (your issue tracker, code host and a database) with least-privilege access. Introduce Claude Projects for shared team knowledge and the Memory tool for long-running agents. Once patterns stabilize, package everything into Plugins for one-step org-wide distribution. Treat the whole setup as code: review it, audit MCP permissions, and keep an enablement owner. A focused Claude enablement workshop can compress this rollout from months to weeks.

Context Engineering for Claude in the Enterprise 2026

How enterprises do context engineering for Claude in 2026: AGENTS.md, Agent Skills, MCP servers, Projects, the Memory tool, Plugins, Routines and Subagents — the shift from prompt engineering to context engineering, with official sources.

Updated: 2 giugno 2026

by Michael Kerkhoff

TL;DR

Context engineering is the 2026 successor to prompt engineering: instead of crafting individual prompts, enterprises architect what Claude knows, can do, and remembers. The core building blocks are AGENTS.md (the repository-level system manifest), Agent Skills (on-demand domain expertise), MCP servers (the tool and data integration layer), Claude Projects (shared workspaces), the Memory tool (durable cross-session context), Plugins, reusable Routines, and Subagents. Treat these as version-controlled "context-as-code" — repeatable, auditable, and team-wide — to roll out Claude reliably across an organization.

Top Picks

AGENTS.md / CLAUDE.md

The foundation of context engineering: a repository-level manifest that gives every Claude session a persistent identity, conventions, and guardrails. Version it in git so the "system prompt" of your AI is reviewable and consistent across the whole team.

Repo-level system prompt — persistent identity, conventions and guardrails for every sessionFree — open convention

Agent SkillsAI-Native

Reusable, filesystem-based packages of domain expertise — workflows, context and best practices that load on demand via progressive disclosure. Skills turn a general agent into a specialist without bloating the context window, and they are portable across Claude apps, Claude Code and the Agent SDK.

On-demand domain expertise — workflows and best practices loaded only when relevantIncluded in Claude / Claude Code

MCP Servers

The Model Context Protocol is the open standard that connects Claude to enterprise data silos and tools (Jira, GitHub, SQL, internal APIs). MCP gives agents a least-privilege, typed integration layer — solving the "last mile" of getting proprietary context into the model.

Standardized connection to enterprise data and tools (Jira, GitHub, SQL, internal APIs)Free — open protocol (Linux Foundation)

Claude ProjectsAI-Native

Persistent workspaces that bundle custom instructions with an uploaded knowledge base for a team or initiative. Projects turn static internal documentation into an "institutional memory" agent that every team member shares, ideal for onboarding and repeatable workflows.

Persistent workspace bundling instructions + a knowledge base for a team or initiativeClaude Pro / Team / Enterprise plans

Memory ToolAI-Native

Durable, cross-session memory so agents retain context without re-onboarding. Combined with context-management/compaction, the Memory tool lets long-running enterprise agents accumulate knowledge over time instead of starting cold every session.

Durable cross-session memory so agents retain context without re-onboardingAPI feature (token-metered)

PluginsAI-Native

Packaged bundles of Skills, slash commands, MCP servers and hooks distributed through marketplaces. Plugins are how an enterprise ships a curated, versioned context setup to every developer in one step — the distribution mechanism for everything below.

Packaged bundles of Skills, commands, MCP servers and hooks for one-step distributionIncluded in Claude Code

Slash Commands & RoutinesAI-Native

Reusable, parameterized prompts and multi-step routines for repeatable workflows — release checklists, code review, incident triage. Stored as markdown in the repo, they standardize "how we ask" across the team and pair with hooks for deterministic automation.

Reusable parameterized prompts and multi-step routines for repeatable workflowsFree / included

SubagentsAI-Native

Isolated sub-agents that offload narrow tasks (research, search, review) and protect the main context window from overflow. Subagents are the orchestration primitive for context engineering — each runs with its own scoped tools and fresh context.

Isolated sub-agents that offload narrow tasks and protect the main context windowIncluded in Claude Code

Comparison Table

Name	What It Solves	Format / Mechanism	Scope / Level	Cost / Availability
1AGENTS.md / CLAUDE.md	Repo-level system prompt — persistent identity, conventions and guardrails for every session	Markdown manifest at repo root (AGENTS.md / CLAUDE.md), git-versioned	Project / repo	Free — open convention
2Agent Skills	On-demand domain expertise — workflows and best practices loaded only when relevant	Filesystem folder: SKILL.md + scripts/resources, progressive disclosure	User / project / org	Included in Claude / Claude Code
3MCP Servers	Standardized connection to enterprise data and tools (Jira, GitHub, SQL, internal APIs)	MCP servers over stdio / Streamable HTTP; TypeScript & Python SDKs	Org / infrastructure	Free — open protocol (Linux Foundation)
4Claude Projects	Persistent workspace bundling instructions + a knowledge base for a team or initiative	Claude (Pro/Team/Enterprise) Projects: custom instructions + knowledge base	Team / workspace	Claude Pro / Team / Enterprise plans
5Memory Tool	Durable cross-session memory so agents retain context without re-onboarding	Memory tool + context management API; file-based memory directory	Agent / session-spanning	API feature (token-metered)
6Plugins	Packaged bundles of Skills, commands, MCP servers and hooks for one-step distribution	Claude Code plugins via marketplaces (.claude-plugin manifest)	Team / org	Included in Claude Code
7Slash Commands & Routines	Reusable parameterized prompts and multi-step routines for repeatable workflows	Markdown slash commands (.claude/commands) + hooks for automation	User / project	Free / included
8Subagents	Isolated sub-agents that offload narrow tasks and protect the main context window	Subagent definitions (.claude/agents) with scoped tools and own context	Task / orchestration	Included in Claude Code

← Scroll horizontally to see all columns

How to Choose

Start with AGENTS.md before anything else: a single repo-level manifest with your conventions, guardrails and identity is the cheapest, highest-leverage context-engineering move. Version it in git so it is reviewable.
Use Skills for "how we do X", MCP for "where the data lives". Skills package repeatable expertise and load on demand; MCP servers give Claude least-privilege, typed access to enterprise systems. Most rollouts need both.
Choose the right scope: AGENTS.md, slash commands and Skills live in the repo (per-project); Projects and the Memory tool persist across sessions for teams; Plugins distribute a curated setup org-wide in one step.
Manage the context window deliberately: prefer Subagents and on-demand Skills over stuffing everything into one long prompt. Engineer high-signal context bundles to cut token cost and reduce drift.
Treat context as code: version, review and audit your AGENTS.md, Skills and MCP configs. For the enterprise, governance, least-privilege MCP access and an audit trail matter as much as raw capability.

Frequently Asked Questions

Related Resources

📖 Related Guides

📝 Related Blog Posts

⚖️ Related Comparisons

🔧 Our Services

Sources & Further Reading

Effective context engineering for AI agents

Anthropic Engineering

Equipping agents for the real world with Agent Skills

Anthropic Engineering

Agent Skills — Overview

Claude API Docs

MCP connector

Claude API Docs

Context Engineering for Coding Agents

martinfowler.com

AGENTS.md — the open standard for agent instructions

AGENTS.md

Context Studios

Pronto per il tuo progetto AI?

Prenota una consulenza gratuita di 30 minuti per discutere le tue esigenze.

Prenota consulenza