Claude Code 2026:
the whole machine.

Check things off as you actually use them — not when you read about them. Reading is trivia; doing is retention. SLEEPER = a high-value thing most people (you, last week) don't know exists. Your progress saves to this browser automatically.

A Skill = SKILL.md (markdown + YAML frontmatter) that loads when you type /name or Claude matches its description. The whole game is progressive disclosure: only descriptions cost context at startup; bodies load on use; reference files load only when read.

A1Skill vs Subagent vs CLAUDE.md rule

Decision rule: if a CLAUDE.md section became a multi-step procedure not a fact, make it a skill. If you keep pasting the same instructions, that's the extraction signal.

A2SKILL.md anatomy + every frontmatter field

---
name: fix-issue
description: Fix a GitHub issue. Use when user says "fix issue", "resolve #N"...
argument-hint: "[issue-number]"
disable-model-invocation: true   # side effects → you fire it, not Claude
allowed-tools: Read Edit Bash(gh *) Bash(git *)
model: sonnet · effort: high · paths: "src/**" · context: fork · agent: Explore
---
# Fix Issue $ARGUMENTS
## Context
!`gh issue view $ARGUMENTS`   # runs BEFORE Claude sees it, output injected
## Steps
1. Read above. 2. Grep code. 3. Fix. 4. Test. 5. git commit -m "fix (#$ARGUMENTS)"

• description is the trigger — capped 1,536 chars, triggers first. Official claude-api skill uses literal TRIGGER when: ... SKIP: ... routing.
• !`cmd` = render-time shell injection (preprocessing, not Claude executing). ${CLAUDE_SKILL_DIR} = portable script paths.
• paths: only auto-activates for matching files. context: fork runs it in a subagent.

A3Progressive disclosure — the directory IS the skill

my-skill/
├── SKILL.md          # <500 lines. Routing + key steps.
├── references/*.md   # deep docs — loaded ONLY when read (informs reasoning)
├── assets/           # templates/fonts — land in the deliverable, never loaded
└── scripts/*.py      # Claude EXECUTES (run --help, do NOT read source)

Budget mechanics: all skill descriptions take ~1% of context at startup. Too many → low-priority ones get silently truncated and stop firing. Diagnose with /doctor; fix with skillOverrides: name-only or raise skillListingBudgetFraction.

A4Steal-worthy official skills

Bookmarks: anthropics/skills (17 official) · travisvn/awesome-claude-skills (12.7k★).

A5Plugins = everything, packaged & shareable

A plugin bundles skills + commands + agents + hooks + MCP + bin/ (auto-PATH) + themes + output-styles, distributed via a marketplace (git repo with marketplace.json). Gives you namespacing (/my-plugin:review) + one-command install + versioning for teammates.

my-plugin/
├── .claude-plugin/plugin.json   # manifest ONLY here
├── skills/ agents/ hooks/ bin/  # everything else at root
└── .mcp.json                    # bundled server, auto-starts
/plugin marketplace add you/marketplace · /plugin install my-plugin@marketplace

Three primitives people conflate: subagents (isolated, return a summary), background agents (full sessions surviving terminal close), agent teams (peers that message each other). Plus claude agents to drive them all.

B1Subagents — the workhorse

---
name: db-reader
description: Read-only DB queries. Do NOT use for writes or schema changes.
model: haiku · tools: Bash · isolation: worktree · memory: project · maxTurns: 20
---
Only execute SELECT. Always LIMIT 100...

• Built-ins: Explore (Haiku search), Plan (read-only research), general-purpose. Agent(worker, researcher) whitelists which sub-subagents may spawn.
• Force one: @"db-reader (agent)" .... Become one: claude --agent db-reader.
• memory: project persists learnings across sessions in .claude/agent-memory/. Gotcha: subagents can't spawn subagents; disk-added agents need a restart.

B2Agent View — claude agents (the new screen, May 11)

Full-screen TUI of every session grouped by state: Needs input, Completed, Working, Ready for review. Summaries auto-refresh every 15s.

Filters: a:name, s:working, #PR. Scriptable: claude agents --json.

B3Agent Teams — peers, not minions

You have CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1. A lead spawns teammates (full sessions) that talk to each other via mailbox + shared task list. Team hooks: TeammateIdle, TaskCreated, TaskCompleted.

B4Background agents & bash

• Agents: claude --bg "..." or /bg; manage with claude logs/attach/respawn/stop/rm; supervisor survives sleep (not shutdown).
• Bash: Ctrl+B backgrounds a running command, Ctrl+T task list, Ctrl+X Ctrl+K kills subagents. 5GB output cap.

B5Worktrees — isolation that makes parallelism safe

• claude --worktree feature-x · subagent isolation: worktree · bg agents auto-isolate before first edit.
• .worktreeinclude auto-copies .env. worktree.symlinkDirectories shares node_modules. Not isolated: DB, ports, fresh node_modules.

CLAUDE.md loads every session, every request — every line is a recurring token tax. Anthropic's own docs: "Bloated CLAUDE.md files cause Claude to ignore your actual instructions." Target <200 lines.

C1The one-question test

For each line: "Would removing this cause a mistake?" If no, cut it. Include: commands Claude can't guess, style that differs from default, repo etiquette, env quirks. Exclude: standard conventions, "write clean code", API docs (link), file-by-file descriptions.

C2Hierarchy, nesting & @-imports

All files concatenate root-down: managed → ~/.claude/CLAUDE.md → project → CLAUDE.local.md. Nested packages/api/CLAUDE.md loads on demand when Claude touches that subtree.

@docs/shared-conventions.md   # import (max 5 hops)
@AGENTS.md                   # share source of truth with other tools
# .claude/rules/api.md → --- paths: ["packages/api/**"] --- loads only on match

C3Three memory systems

/memory = audit panel (what's loaded). # = quick-save. /doctor = budget check.

The enforcement layer. CLAUDE.md is a request; a hook is a guarantee. MCP plugs in tools. Sandbox + permissions decide what any of it can touch.

D1Hooks — ~30 events, deterministic automation

"hooks":{"PostToolUse":[{"matcher":"Edit|Write",
  "hooks":[{"type":"command","if":"Edit(src/**)",
    "command":"jq -r '.tool_input.file_path' | xargs npx prettier --write"}]}]}

Types: command, 💡 http, mcp_tool, prompt (LLM-judged). Exit 0=ok, 2=block+feed stderr to Claude. if scopes to specific calls.

D2MCP — and Tool Search (why your context survives)

claude mcp add --transport http sentry https://mcp.sentry.dev/mcp
claude mcp add --transport stdio playwright -- npx -y @playwright/mcp@latest
# OAuth: just run /mcp and follow the browser login

Scopes: local / project (.mcp.json) / user. Resources: @server:type://path. Prompts: /mcp__server__prompt. 💡 Tool Search on by default (ENABLE_TOOL_SEARCH) — only tool names load at start, definitions on demand. This is why your Linear+Supabase+Chrome+Gmail roster doesn't drown context. alwaysLoad: true pins a server. Sweet spot: GitHub + Context7 + Playwright.

D3Sandbox · Permissions · Output styles

Other knobs: effortLevel: xhigh, alwaysThinkingEnabled, prompt-caching env vars, 💡spinnerTips/spinnerVerbs (custom loading text), attribution, apiKeyHelper.

The platform grew a cloud execution surface, scheduled agents, self-driving loops, and phone control in four months. Reverse-chronological, complete.

Mayv2.1.139–146

• Agent View (claude agents) — the dashboard. /goal 💡 — set a completion condition, Claude auto-runs turns until it holds. Fast mode → Opus 4.7. Rewind "summarize up to here". claude agents --json.

Aprv2.1.92–136

• Opus 4.7 (xhigh effort, your default). Routines 💡 — cloud agents on schedule/GitHub-event, machine off. Mobile push notifications. /usage (merged /cost+/stats). Ultraplan — plan mode in the cloud. Monitor tool 💡 — bg watchers streaming into the convo. /loop — self-pacing scheduled tasks. /autofix-pr — watches CI, pushes fixes till green. /ultrareview — cloud bug-hunt fleet. Plugins from .zip/--plugin-url. Cross-project Ctrl+R. claude project purge. Resume by PR URL. Windows without Git Bash.

Marv2.1.83–91

• Computer Use 💡 (CLI + Desktop) — Claude clicks through GUIs to verify its own work. Auto Mode classifier permissions. PR auto-fix on web. Transcript search (/ in transcript). Conditional hooks (if). Flicker-free render (CLAUDE_CODE_NO_FLICKER=1). /powerup 💡 — in-terminal interactive lessons. Plugin bin/ on PATH.

Febv2.1.32–82

• Remote Control 💡 — drive a local session from your phone via QR (outbound HTTPS only). Agent Teams (experimental). Opus 4.6 + Fast mode. /loop launch + Linux PID-namespace sandboxing.

The built-in commands worth knowing. Type / in any session to see yours (skills + plugins add more).

CLI (no slash): claude agents · claude --bg "..." · claude --worktree x · claude --agent X · claude mcp add · claude plugin validate · claude logs/attach/respawn/stop · claude project purge.