Claude Prompt Builder

Release Notes — All Branches As of Feb 25, 2026 · API Feb 19 · CC v2.1.55

Current Models — Active

Claude Opus 4.6 LATEST FLAGSHIP

claude-opus-4-6

All plans

Feb 5, 2026

Claude Sonnet 4.6 LATEST BALANCED

claude-sonnet-4-6

All plans

Feb 17, 2026

Claude Opus 4.5

claude-opus-4-5-20251115

Nov 24, 2025

Claude Sonnet 4.5

claude-sonnet-4-5-20250929

Sep 29, 2025

Claude Haiku 4.5

claude-haiku-4-5-20251001

Oct 15, 2025

Claude Sonnet 4 / Opus 4.1

claude-sonnet-4-20250514 · claude-opus-4-1-20250805

Removed from selector

May–Aug 2025

⛔ Retired — All requests return error

Sonnet 3.7 · Haiku 3.5

claude-3-7-sonnet-20250219 · claude-3-5-haiku-20241022

Retired Feb 19, 2026

✗

Claude Opus 3 · Claude 2.x · Sonnet 3.5

Various — all retired

Retired 2025

✗

Claude Haiku 3 ⚠ Deprecation Notice

claude-3-haiku-20240307

Retiring Apr 19, 2026

⚠

Feature Compatibility by Model

Feature	Opus 4.6	Sonnet 4.6	Opus 4.5	Sonnet 4.5	Haiku 4.5
Adaptive thinking	✓ auto	✓ auto	—	—	—
Effort param (GA)	low/med/high/max	low/med/high	low/med/high	—	—
Extended thinking (manual budget_tokens)	deprecated	✓	✓	✓	—
Interleaved thinking	✓ auto via adaptive	beta header reqd	beta header reqd	—	—
Prefill assistant messages	✗ returns 400	✓	✓	✓	✓
128K output tokens	✓	64K	32K	16K	16K
1M context window	beta (Tier 4)	beta (Tier 4)	beta (Tier 4)	beta (Tier 4)	—
Fast mode	✓ research preview	—	—	—	—
Compaction API	✓ (beta)	—	—	—	—
Data residency (inference_geo)	✓	✓	—	—	—
Dynamic web filtering	✓	✓	—	—	—
Structured outputs (GA on 1P)	✓	✓	✓	✓	✓
Web search / Code execution	✓	✓	✓	✓	✓
Agent Skills (API beta header)	✓	✓	✓	✓	✓

✓ auto = auto-enabled beta = requires beta header or special access ✗ returns 400 = breaking change

✓ GA — API / Developer Platform

Adaptive Thinking All plans Opus 4.6 · Sonnet 4.6

Claude decides when and how much to think. Replaces manual budget_tokens. Interleaved thinking is auto-enabled. Effort: low / medium / high (default) / max.

API: thinking={"type": "adaptive"}, effort="high" CC settings.json: alwaysThinkingEnabled: true CC toggle: Option+T (Mac) / Alt+T (Linux/Win) CC command: /effort low|medium|high|max

Effort Parameter All plans No beta header

4 levels: low · medium · high (default) · max. Replaces budget_tokens. Works on Opus 4.6, Sonnet 4.6.

API: effort="low"|"medium"|"high"|"max" # no beta header CC: /effort max

128K Output Tokens Opus 4.6 only

Doubled from 64K. Sonnet 4.6 remains 64K. SDK requires streaming for large max_tokens.

Structured Outputs All API plans Beta on Bedrock/Foundry

GA on Anthropic API (Sonnet 4.5, Opus 4.5, Haiku 4.5, Opus 4.6). Still beta on Amazon Bedrock and Microsoft Foundry. Note: output_format moved to output_config.format.

API: output_config={"format": {"type": "json_schema", "schema": {...}}} # Old output_format is deprecated but still works

Web Search + Web Fetch + Code Execution All API plans

All 3 tools now GA — no beta header. Code execution is free when paired with web search or web fetch. Dynamic filtering (beta) uses code execution to pre-filter search results.

Memory Tool + Tool Search + Programmatic Tool Calling GA

All three now GA — no beta header. Previously required beta headers.

Fine-Grained Tool Streaming All models · All platforms

GA on all models and platforms. No beta header required.

Automatic Prompt Caching All API plans Preview on Azure

Single cache_control field auto-manages cache points as conversation grows. No manual breakpoints needed. Works alongside existing block-level cache control.

API: Add cache_control to request body # Claude API + Azure (preview)

Data Residency Controls All API plans Opus 4.6+ only

Specify inference region per request. US-only = 1.1× pricing. Only applies to models released after Feb 1, 2026.

API: inference_geo="global"|"us" # 1.1x for "us"

Agent Skills All plans on claude.ai API: beta header

GA on claude.ai with extended thinking. API still uses beta header skills-2025-10-02. Pre-built skills: PPTX, XLSX, DOCX, PDF. Org-wide management on Team/Enterprise.

API beta header: anthropic-beta: skills-2025-10-02

✓ GA — Claude Code (v2.1.55)

Git Worktrees Pro / Max / Team / Enterprise

Each worktree = isolated branch + directory. No file conflicts between parallel sessions. WorktreeCreate/WorktreeRemove hook events available.

CLI: claude --worktree feature-name # or -w Agent frontmatter: isolation: worktree Hooks: WorktreeCreate, WorktreeRemove events

Custom Agents + Background Agents Pro / Max / Team / Enterprise

Define agents in .claude/agents/*.md. Background agents run without blocking main thread. Kill all: Ctrl+F.

Agent frontmatter: background: true List agents: claude agents Kill all bg agents: Ctrl+F (two-press confirm)

Auto Memory All CC plans

Claude automatically records and recalls memories as it works. Persists across sessions in the project directory. Memory scope can be set per-agent (user / project / local).

Disable: CLAUDE_CODE_DISABLE_AUTO_MEMORY=1 Agent frontmatter: memory: user|project|local

CLAUDE.md from --add-dir Paths All CC plans

By default Claude Code only loads CLAUDE.md from the working directory. This flag enables loading CLAUDE.md from directories added via --add-dir (or /add-dir) as well.

settings.json env: CLAUDE_CODE_ADDITIONAL_DIRECTORIES_CLAUDE_MD=1

Plugin System All CC plans

Extend Claude Code with custom commands, agents, hooks, and MCP servers. Plugins auto-load hooks from ~/.claude-plugin/hooks/hooks.json — do NOT declare a hooks field in plugin.json (causes duplicate errors in v2.1+). Cowork plugins with admin controls available for Team/Enterprise (Feb 24).

Install: /plugin install <source> Manage: /plugin enable|disable|marketplace|validate Dir: ~/.claude-plugin/ (or .claude-plugin/ per-project)

Subagent + Agent Restrictions All CC plans

Restrict which sub-agent types a custom agent can spawn. Background agents run without blocking main thread; kill all with Ctrl+F (two-press). List configured agents via CLI.

Agent "tools" frontmatter: Task(agent_type) # restrict spawnable agent types Agent frontmatter: background: true # always run as background Kill all bg agents: Ctrl+F (two-press confirm) List agents: claude agents

MCP Server Configuration All CC plans

Per-session MCP toggles via @-mention or /mcp. Disable noisy servers in settings.json. Timeout controls for slow servers or large file uploads. Pre-configured OAuth since v2.1.30.

Disable specific servers: "disabledMcpServers": ["server1", "server2"] Connection timeout: MCP_TIMEOUT=30000 # ms, server start Tool timeout: MCP_TOOL_TIMEOUT=60000 # ms, single tool call Toggle in session: @mention server name, or /mcp

Session Language + Misc Settings All CC plans

Configure Claude Code's response language. Skill hot-reload: skills created or modified in ~/.claude/skills/ are immediately available without restarting. Task system rollback env var if needed.

Response language: "language": "japanese" # in settings.json Skill hot-reload: automatic — no restart needed Skill fork context: context: fork # in skill frontmatter Task system rollback: CLAUDE_CODE_ENABLE_TASKS=false

Remote Control All CC plans

Enables local environment serving for external builds. Added in v2.1.51.

CLI: claude remote-control

ConfigChange Hook Enterprise auditing

Fires when configuration files change during a session. Enables security auditing and optional blocking of settings changes. disableAllHooks now respects managed settings hierarchy.

Simple Mode All plans

Strips Claude Code to minimum: Bash tool + file edit only. Disables skills, session memory, custom agents, CLAUDE.md, MCP tools, attachments, hooks.

Env var: CLAUDE_CODE_SIMPLE=1

✓ GA — Claude.ai Product

Memory All paid plans

Claude remembers context across chats. Incognito mode available to exclude sessions. Import/export supported.

Claude in Excel + Claude in PowerPoint Paid plans (Max/Team/Enterprise for Excel)

Excel: Opus 4.6 powered, native pivot tables, conditional formatting, charts. PowerPoint: Add-in side panel, direct presentation editing. Both released/updated Feb 5.

Self-Serve Enterprise Enterprise

Purchase Enterprise plans directly — no sales conversation required. Single seat type: Claude + Claude Code + Cowork. Available Feb 12.

Analytics API Enterprise only

Programmatic access to Claude and Claude Code usage data. Per-organization, per-day aggregation.

Health Data on Mobile Pro / Max · US only

Claude reads activity, sleep, workouts from iOS Health and Android Health Connect. Android 14+ required.

Claude Code on Team Plan Standard Seats Team

Claude Code access now included with every Team plan standard seat. No premium upgrade needed. Added Jan 16.

⚗ Experimental / Research Preview / Beta

Agent Teams (Claude Code) Research Preview All CC plans

Multi-agent orchestration. Lead + teammates each with own context window, shared task list, and inbox. Delegate mode restricts lead to coordination-only (spawning, messaging, task list — no Edit/Write/Bash). Known bug: delegate mode has tool-access-inheritance issue with teammates (track issue #25037). No session resumption for in-process teammates. One team per session — no nested teams.

Enable: "env": { "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1" } # in settings.json Pane mode: "teammateMode": "auto"|"in-process"|"tmux" # auto = smart detect tmux mode: requires tmux installed (brew install tmux) iTerm2 mode: tmux -CC (native pane mapping, suggested entrypoint) VS Code: split-pane mode NOT supported — use in-process Delegate mode: Shift+Tab mid-session # toggle coordination-only Keyboard: Shift+Down (cycle teammates), Ctrl+T (task list), Enter (view session), Esc (interrupt) TeammateIdle hook: exit code 2 = send stderr as feedback, teammate keeps working TaskCompleted hook: fires when any teammate marks a task done Agent frontmatter: Task(agent_type) in "tools" # restrict spawnable sub-agents Memory scope: memory: user|project|local # in agent frontmatter

Fast Mode Research Preview CC: all plans + extra usage API: waitlist

2.5× faster output at 6× cost ($30/$150 per MTok). Same model intelligence. Opus 4.6 only. NOT on Bedrock / Vertex / Foundry. CC: auto-switches to Opus 4.6 when enabled. Includes 1M context at no extra charge.

CC settings.json: "fastMode": true CC toggle: /fast (persists across sessions) CC requires: extra usage enabled in account settings API beta header: anthropic-beta: fast-mode-2026-02-01 API param: speed="fast" API: waitlist at claude.com/fast-mode

1M Token Context Window Beta API: Tier 4 / custom CC: all plans (Opus 4.6)

Opus 4.6 and Sonnet 4.6. Long context pricing above 200K tokens (stacks with other pricing modifiers). Fast Mode includes 1M context at no extra charge. CC: enabled by default for Opus 4.6, disable if needed.

CC disable: CLAUDE_CODE_DISABLE_1M_CONTEXT=1 API: Requires API Tier 4 or custom rate limit agreement

Compaction API Beta All API plans · Opus 4.6 only

Server-side context summarization for effectively infinite conversations. Claude Code has built-in auto-compaction that triggers near context limit. Manual trigger via /compact. Only available on Opus 4.6 via API; Sonnet 4.6 use client-side compaction via SDK instead.

API: Enabled via API configuration — check current docs for beta header CC command: /compact (manual trigger at any time) CC auto: triggers when approaching context limit SDK alt: client-side compaction via tool_runner (Python/TS SDKs)

Cowork Research Preview Max full · Pro limited

Claude Code's agentic capabilities for knowledge work. Runs locally in isolated VM, accesses local files and MCP integrations. macOS desktop only. Max 20× for full access. Pro: limited preview since Jan 16.

Requires: Claude Desktop app (macOS only) Access: Max plan (full) or Pro plan (limited)

Interleaved Thinking Public Beta Sonnet 4.5 · Opus 4.5 · Sonnet 4.6

Claude thinks between tool calls for better multi-tool reasoning. On Opus 4.6, automatically enabled via adaptive thinking — remove the beta header. On Sonnet 4.6, Opus 4.5, and Sonnet 4.5: still requires the beta header with manual extended thinking.

Beta header: anthropic-beta: interleaved-thinking-2025-05-14 Opus 4.6: REMOVE header — auto-enabled via adaptive thinking Sonnet 4.6: still requires beta header + thinking: {type: "enabled", budget_tokens: N}

Files API Public Beta All API plans

Upload files once, reference them in multiple Messages API calls or code execution. Avoids re-uploading large files every request. Introduced May 2025 with Claude 4 / Sonnet 4.

Upload: POST /v1/files Reference: {type: "document", source: {type: "file", file_id: "file_..."}} Docs: platform.claude.com/docs/en/build-with-claude/files

MCP Connector via API Public Beta All API plans

Connect to remote MCP servers directly from the Messages API without running a local proxy. Pass mcp_servers in the request. Introduced May 2025. Wildcard tool permissions: mcp__server__*.

API param: mcp_servers: [{type: "url", url: "https://...", name: "..."}] Beta header: anthropic-beta: mcp-client-2025-04-04 Wildcard permission: mcp__server__* (grants all tools from a server)

Context Editing / Thinking Block Clearing Beta All API plans

Automatically clears older thinking blocks when approaching token limits, preserving text blocks. Prevents context windows bloating from large reasoning traces in long agentic sessions. Client-side compaction via SDK tool_runner also available.

Beta header: anthropic-beta: clear_thinking_20251015 SDK (Python/TS): use tool_runner with auto client-side compaction CC: auto-triggers near context limit · /compact for manual trigger

Structured Outputs on Bedrock / Foundry Public Beta

GA on Anthropic API (1P) but still in public beta on Amazon Bedrock and Microsoft Azure Foundry. Requires beta header on those platforms only.

Bedrock/Foundry header: anthropic-beta: structured-outputs-2025-11-13 Anthropic 1P: GA — no header needed

⚠ Deprecated / Breaking Changes

Prefill on Opus 4.6 Breaking — Returns 400

Prefilling assistant messages ({"role":"assistant","content":"..."}) is not supported on Opus 4.6 and returns a 400 error. Still works on Sonnet 4.6, Opus 4.5, Sonnet 4.5, Haiku 4.5. Alternatives: structured outputs, system prompt, output_config.format.

thinking type "enabled" + budget_tokens on Opus 4.6 Deprecated (Opus 4.6 only)

Still functional but deprecated on Opus 4.6. Will be removed in a future release. Migrate to thinking: {type: "adaptive"} + effort parameter. Still valid on Sonnet 4.6, Opus 4.5, Sonnet 4.5.

interleaved-thinking-2025-05-14 header on Opus 4.6 Deprecated (Opus 4.6 only)

On Opus 4.6: safely remove — adaptive thinking auto-enables interleaved thinking. On Sonnet 4.6, Opus 4.5, Sonnet 4.5: still required. Leaving it in Opus 4.6 requests is harmless but unnecessary.

output_format parameter Deprecated — All models

Moved to output_config.format. Old parameter still functional during transition but will be removed in a future release.

Before: output_format={"type": "json_schema", ...} After: output_config={"format": {"type": "json_schema", ...}}

Legacy beta headers — remove from all requests Legacy

These headers are no longer required and should be removed from API requests to avoid confusion.

Remove: anthropic-beta: token-efficient-tools-2025-02-19 Remove: anthropic-beta: output-128k-2025-02-19 # Both are legacy — functionality is now standard GA

temperature + top_p together Removed — Opus 4.1+

Opus 4.1 and newer: specifying both temperature and top_p is not allowed. Use only one. Applies to Opus 4.1, Opus 4.5, Opus 4.6.

Claude Opus 4 / Opus 4.1 — Removed from model selector Removed Jan 2026

Removed from Claude and Claude Code model selector in January 2026. Anthropic recommends migrating to Opus 4.5 or Opus 4.6. API strings may still work during transition period but not officially supported.

Sources (verified Feb 25, 2026): Opus 4.6 blog API release notes What's new in 4.6 Product release notes CC CHANGELOG.md CC releases

Import

Profile

Powers

Generate

◈ AI-Powered Import

Send your project docs to Opus with the prompt below. Paste the JSON it returns to auto-fill all fields.

Extraction Prompt

Paste Opus JSON output

Select Profile

Project Identity

project_name short identifier

goal one sentence

Stack

BACKEND

FRONTEND

DATABASE

RUNTIME

Approval Mode

PLAN_FIRST safest

Architecture plan → approval → build

BUILD_DIRECT fastest

Build immediately, no plan needed

MVP_FIRST exploratory

Smallest version first, extend after review

⚡ Power Features

Each toggle appends a specialized XML block to your prompt. When ON, the required config/launch params are shown below the toggle.

Multi-Agent & Parallelism

Agent Teams Research Preview All CC plans

Lead agent orchestrates teammates, each in their own context window. Shared task list and inbox messaging. Split-pane visualization in tmux/iTerm2.

Required setup before launch: settings.json env section: "env": { "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1" } For tmux split panes (recommended): "teammateMode": "tmux" Install tmux: brew install tmux Launch Claude from inside a tmux session For iTerm2 native panes: "teammateMode": "iterm2" Then start with: tmux -CC Delegate mode (lead = coordination only): Shift+Tab mid-session to toggle

Git Worktrees Stable · All CC plans

Each agent or feature gets isolated branch + directory. No file conflicts between parallel sessions.

Launch: claude --worktree feature-name (or -w) Agent frontmatter: isolation: worktree Cleanup after merge: git worktree remove [path]

Background Agents Stable · All CC plans

Agents run without blocking the main thread. Main thread continues independently. Kill all with Ctrl+F.

Agent frontmatter: background: true Kill all: Ctrl+F (two-press confirmation) List agents: claude agents

Reasoning & Quality

Fast Mode Research Preview Extra usage required

2.5× faster output at 6× token cost ($30/$150 per MTok). Same Opus 4.6 intelligence. Includes 1M context at no extra charge. Not available on Bedrock/Vertex/Foundry.

CC settings.json: "fastMode": true CC toggle: /fast (per session) Requires: extra usage enabled in Claude account settings Note: Auto-switches to Opus 4.6 when enabled API: Waitlist at claude.com/fast-mode betas=["fast-mode-2026-02-01"], speed="fast"

Adaptive Thinking Stable · Opus 4.6 default

Explicitly instruct Claude to use adaptive thinking for architecture decisions and complex tradeoffs. Use effort=max for highest capability.

CC settings.json: "alwaysThinkingEnabled": true CC toggle: Option+T (Mac) / Alt+T (Linux/Win) CC effort: /effort max API: thinking={"type": "adaptive"}, effort="max"

Builder-Validator Chain Stable

One agent builds, a second reviews independently before any commit. Catches regressions and logic errors before they land.

Best paired with: Agent Teams or Background Agents Cost note: Approximately doubles agent usage

Session Management

Context Budget Mode Stable

Auto-compact at 70% usage. Save session state before limit. Never stop mid-task regardless of remaining budget.

CC command: /compact (manual trigger, also auto at ~98%) API beta: anthropic-beta: compact-2026-01-12 Opus 4.6 only on API. CC supports all models.

CLAUDE.md Self-Update Stable

After every correction or PR, Claude updates its own rules file so mistakes don't recur across sessions.

Init: /init (creates CLAUDE.md if missing) Auto memory: CC records memories automatically across sessions Best practice: Keep CLAUDE.md under ~100 lines (2,500 tokens)

✓ Prompt Ready

Copy from the right panel and paste into a new Claude Code or claude.ai session to launch your project.

Mid-Session Commands

SWITCH MODE: PLAN_FIRST

SWITCH MODE: BUILD_DIRECT

SWITCH MODE: MVP_FIRST

EXPLAIN THIS

SCOPE LOCK

EXTEND IT

HARD STOP

Claude Prompt Builder

Prompt will appear here