Claude Code Extension Points: When to Use What¶

Choose the right extension point — CLAUDE.md, rules, skills, hooks, subagents, MCP servers, or plugins — based on enforcement needs, context cost, and portability.

At a glance¶

Extension point	Purpose	Loads when	Context cost
CLAUDE.md	Project conventions	Session start	Full content every turn
`.claude/rules/`	Path-targeted rules	Session start or file match	Path-scoped
Skills	Workflows and reference knowledge	Descriptions at start; full on invoke	Progressive
Hooks	Deterministic shell commands at lifecycle events	On trigger	Zero (external)
MCP servers	External tool and service access	Session start	Schemas only
Subagents	Isolated agents for focused tasks	On spawn	Zero (isolated)
Agent teams	Multi-agent coordination	On session creation	Zero per teammate
Plugins	Distribution bundles	On install	Varies

Decision framework¶

flowchart TD
    A[Need to extend Claude Code] --> B{Must happen 100%<br/>of the time?}
    B -- Yes --> C[Hook]
    B -- No --> D{Needs external<br/>service access?}
    D -- Yes --> E[MCP server]
    D -- No --> F{Applies to every<br/>session universally?}
    F -- Yes --> G{Short and<br/>global?}
    G -- Yes --> H[CLAUDE.md]
    G -- No --> I[.claude/rules/]
    F -- No --> J{Repeatable workflow<br/>or reference knowledge?}
    J -- Yes --> K[Skill]
    J -- No --> L{Needs isolated context<br/>or parallel execution?}
    L -- Yes --> M[Subagent]
    L -- No --> N{Multiple agents<br/>need to coordinate?}
    N -- Yes --> O[Agent team]
    N -- No --> P{Needs to be shared<br/>across repos/teams?}
    P -- Yes --> Q[Plugin]
    P -- No --> H

Based on the official features overview and community models (Ottmann, GenAI Unplugged).

CLAUDE.md vs .claude/rules/ vs skills¶

All three deliver instructions, but differ in scope and context cost.

	CLAUDE.md	.claude/rules/	Skills
When to use	Core conventions, under ~200 lines	Domain rules scoped to file paths	On-demand reference or `/name` workflows
Loads	Every session, unconditionally	Session start or on file-path match	Descriptions at start; full on invocation
Context cost	High — always present	Medium — path-targeted	Low — progressive
Typical content	Architecture, test commands, conventions	Lint rules for `frontend/`, API rules for `api/`	Procedures, checklists, templates

Keep CLAUDE.md lean: path-specific rules go in .claude/rules/, detailed procedures go in skills. See Hierarchical CLAUDE.md and Progressive Disclosure.

Deterministic vs probabilistic¶

Deterministic — the model cannot override:

Hooks fire at 18 lifecycle events. The agent cannot skip or override them.
CLAUDE.md loads unconditionally at session start (memory docs).

Probabilistic — the model decides when to invoke:

Skills: Claude invokes them based on description relevance (skills docs). Set disable-model-invocation: true for explicit-only.
Subagent delegation: the parent decides when to spawn.

For non-negotiable rules, prefer hooks. See Hooks vs Prompts.

When extension points combine¶

CLAUDE.md and hooks: CLAUDE.md states the rule; a PreToolUse hook enforces it. See Hooks for Enforcement vs Prompts for Guidance.
Skills and subagents: a skill defines the procedure; a subagent runs it in isolation.
MCP and skills: MCP exposes external tools; a skill provides the workflow that uses them.
Plugins bundle agents, skills, hooks, and MCP configs for distribution. Plugins solve distribution, not logic.

Commands are merged into skills — existing .claude/commands/ files continue to work.

Why the determinism boundary matters¶

The CLI spawns hooks as shell processes at lifecycle events, independent of the model's token stream. The model cannot see or suppress them. Skills, subagents, and .claude/rules/ are text in the model's context, so a confusing context can cause the model to skip them. Anything routed through model reasoning is probabilistic; anything run at the infrastructure layer is deterministic.

CLAUDE.md's high context cost follows from this: it is injected into every request unconditionally. Skills load full content only on invocation.

When this backfires¶

Overlapping extension points: a path-scoped rule that also needs enforcement requires both a .claude/rules/ entry and a hook. You then maintain the rule in two places, which creates drift.
Hook proliferation: hooks applied to stylistic preferences, rather than non-negotiable compliance, add startup latency and more failure modes.
CLAUDE.md bloat: at 500+ lines, unconditional injection weakens instruction-following on unrelated tasks. Move domain content to rules or skills early.
Premature plugins: bundling before two or more repos need the config adds release overhead. A shared git subtree is cheaper at small scale.

Security: per-server MCP trust¶

Before v2.1.69, .mcp.json silently trusted all servers. Claude Code now shows a per-server trust dialog on first session (changelog) — automated setups relying on silent enablement will see a prompt per server after updating.

Deprecation: /output-style → /config¶

/output-style was deprecated in v2.1.73 for /config (changelog). Output style is now fixed at session start to preserve prompt cache hit rates — mid-session changes invalidated the cache. Custom style directories (~/.claude/output-styles/, .claude/output-styles/) still work. See System Prompt Replacement for substituting a domain-specific identity for the default persona.

Example¶

A team wants to enforce that all SQL migrations include a DOWN migration. They also want a reusable database review workflow and access to their internal schema registry.

Requirement	Extension point	Why
Every migration file must contain `DOWN`	Hook (`PreToolUse` on `Write`)	Non-negotiable — must fire 100% of the time, the model cannot skip it
"Review this migration" workflow	Skill (`.claude/skills/review-migration.md`)	On-demand procedure with steps; loads only when invoked
"All SQL uses snake_case" convention	`.claude/rules/db.md` with glob `db/migrations/**`	Path-scoped rule; only loads when touching migration files
Schema registry lookup	MCP server	External service access; exposes `get_schema` tool
Share all of the above across repos	Plugin	Bundles the hook, skill, rule, and MCP config for `npm install` distribution

The hook script (hooks/check-down-migration.sh) runs deterministically. The skill and rule are probabilistic but scoped. The MCP server bridges to an external system. The plugin packages everything for distribution.

Key Takeaways¶

Start with the decision flowchart: enforcement need, service access, scope, and portability each point to a different extension point
CLAUDE.md is for short, universal conventions; .claude/rules/ for path-targeted rules; skills for on-demand knowledge
Hooks are the only fully deterministic extension point — use them when compliance is non-negotiable
Skills load progressively; hooks and subagents have zero context cost — choose based on context budget
Plugins are distribution packaging, not a separate logic layer
MCP servers now require per-server trust approval (v2.1.69) — silent bulk enablement is no longer supported
/output-style is deprecated in favor of /config (v2.1.73) — output style is fixed at session start for cache efficiency