liam-jons Claude Capabilities Awareness
Last updated: 2026-03-20 | Models through: Sonnet 4.6 | Claude Code: v2.1.80+
Rules
Follow these rules when answering questions about Claude's capabilities:
- Never guess availability. If a feature's platform support is not listed below, read the relevant reference file before answering.
- Never state prices. Defer to the pricing page (
anthropic.com/pricing). Pricing changes frequently and varies by context length and model. - Never claim Claude can do embeddings or fine-tuning. Recommend Voyage for embeddings. For behaviour customisation, recommend system prompts, skills, or Projects.
- Distinguish platforms. Claude.ai, Desktop, Claude Code, CoWork, and the API have different feature sets. Always scope your answer to the user's platform. Use the Platform Matrix below.
- Flag deprecations. When a user references
budget_tokens,output_format, prefill (Opus 4.6), or deprecated models (Sonnet 3.7, Haiku 3.5), tell them the replacement. See Breaking Changes below. - Use reference files for code examples. This file has facts and decisions. Code samples and detailed configs live in
references/. - Match the extension to the problem. When users describe workflows, recommend the right extension pattern from the Extension Patterns table. Do not default to "use MCP" or "use a skill" without checking fit.
Current Models
| Model | Context | Max Output | Thinking | Key Trait |
|---|---|---|---|---|
| Opus 4.6 | 1M native | 128K | Adaptive | Most capable, complex reasoning |
| Sonnet 4.6 | 1M native | 64K | Adaptive | Balanced speed/capability |
| Haiku 4.5 | 200K | 64K | Extended | Fast, high-volume tasks |
Legacy: Sonnet 4.5, Opus 4.5 still available. Sonnet 4.5/4 access 1M via beta header context-1m-2025-08-07 (tier 4+). Opus 4.6 and Sonnet 4.6 need no header for 1M.
Read references/model-specifics.md when the user needs model IDs, provider-specific details, or migration guidance.
Platform Matrix
When the user asks "Can Claude do X?", check this table first. Answer with the specific platform context.
| Capability | Claude.ai | Desktop | Claude Code | CoWork |
|---|---|---|---|---|
| Skills (auto-invoke) | Yes (ZIP) | Yes (ZIP) | Yes (filesystem) | Yes |
| Skills (slash /name) | -- | -- | Yes | Via plugins |
| MCP | Connectors | Settings | Full (stdio/HTTP/SSE) | Via plugins |
| Projects | Yes | Yes | -- (use CLAUDE.md) | -- |
| Plugins/Hooks | -- | -- | Yes | Plugins only |
| Subagents/Teams | -- | -- | Yes | Sub-agents |
| Background/Loop | -- | -- | Yes | Long-running tasks |
| Cross-conv. memory | Projects | Projects | CLAUDE.md + skills | Instructions |
| Code Review | -- | -- | Managed (Teams/Ent) | -- |
| Web sessions | claude.ai/code | -- | --remote / /teleport |
-- |
| Remote Control | View/steer | -- | Host session | -- |
| Slack integration | -- | -- | @Claude -> web session | -- |
| Channels | -- | -- | Research preview (v2.1.80) | -- |
| Dispatch (mobile) | -- | -- | -- | Yes (Pro/Max) |
| Scheduled tasks | -- | -- | -- | Yes |
| File outputs | -- | -- | Yes | Excel, PPT, docs |
Platform-Specific Guidance
When the user is on Claude.ai/Desktop: Install skills as ZIP via Settings > Capabilities > Skills. Auto-invocation from natural language (no slash commands). Use Projects for persistent context. MCP Apps for interactive UIs.
When the user is on Claude Code: Full extension system: skills (5 bundled skills: /simplify, /batch, /debug, /loop, /claude-api), plugins, hooks (shell + HTTP), subagents, agent teams, MCP, CLAUDE.md + .claude/rules/. Key features: background tasks (Ctrl+B), /loop scheduling, cron tools, /btw (side questions — forks context, no tools, single-turn, never enters history), /effort, /branch, worktrees (--worktree/-w). Code Review (managed, Teams/Ent only — @claude review, REVIEW.md config). Remote Control (claude remote-control or --remote-control — Pro, Max, Team, Enterprise; mobile QR access). Web sessions (--remote / /teleport). Slack (@Claude -> auto Code sessions). Read references/claude-code-specifics.md for details.
When the user is on CoWork: Autonomous background agent with local file access. Professional outputs (Excel with formulas, PowerPoint, formatted docs). Dispatch from phone (Pro/Max only — single persistent thread, desktop must be awake). Scheduled tasks (on-demand or automated cadence). Plugins bundle skills + connectors + sub-agents. Requires Desktop app (macOS Universal, Windows x64). Paid plans only. Safety: Be selective about file access (dedicated working folder), monitor scheduled tasks, limit web access to trusted sites, evaluate plugins carefully. Limitations: No cross-session memory, no audit logging, local conversation storage only, desktop app must stay open. Read references/claude-code-specifics.md Desktop App section for details.
When the user wants multi-step orchestration from Claude.ai/Desktop: They need either Claude Code (full agent) or the Agent SDK (Python/TypeScript) or custom orchestration via the Messages API with tool use. Claude.ai and Desktop do not support code execution, filesystem access, or multi-step workflows.
API (direct, Bedrock, Vertex AI, Azure): All model capabilities. Bedrock/Vertex may lag on newest features. Model IDs differ by provider. Read references/model-specifics.md for details.
Extension Patterns
When a user describes a need, recommend the right pattern:
| Signal | Recommend | Platform |
|---|---|---|
| "Always do X" rules | CLAUDE.md (~200 lines, .claude/rules/ overflow) |
Claude Code |
| Persistent project context | Projects | Claude.ai, Desktop |
| Reusable knowledge/workflow | Skill | All platforms |
| External service connection | MCP | All (varies) |
| Focused task isolation | Subagent | Claude Code |
| Parallel coordination | Agent team | Claude Code |
| "After every edit/commit" | Hook (deterministic, no tokens) | Claude Code |
| "Check every X minutes" | /loop or Cron | Claude Code |
| Bundle + distribute | Plugin | Claude Code |
| "I always...", "every time..." | Skill (capture as workflow) | All |
| "Share this setup with team" | Plugin | Claude Code |
| "How do I get Claude to do X?" | Check skills.sh first | All |
Key distinctions: CLAUDE.md = "always know this". Projects = persistent context (Claude.ai/Desktop). Skill = "know this when relevant" (all platforms). Skills are content, subagents are isolated workers. MCP gives the ability to act, a skill teaches how to act well.
skills.sh: Public registry. npx skills add (Claude Code) or download ZIP (Claude.ai/Desktop). find-skills meta-skill searches the registry.
Core Capabilities
Vision: All models analyze images (JPEG, PNG, GIF, WebP) via base64, URL, or file_id. Multi-image supported.
PDFs: 32MB max, 600 pages per request (100 for 200K-context models). Images and PDF pages share the 600 limit. Via URL, base64, or Files API. Bedrock requires citations enabled for visual PDF analysis.
Streaming: stream: true for SSE. Fine-grained tool streaming (GA) progressively streams tool parameters.
Thinking: Adaptive thinking (thinking: {type: "adaptive"}) for Opus/Sonnet 4.6. Effort parameter (GA, all models: low/medium/high). budget_tokens deprecated on 4.6 models. Fast mode (research preview, Opus 4.6 only, speed: "fast" + header, significantly higher cost).
Context: 1M native on Opus/Sonnet 4.6. Compaction API for infinite conversations. Prompt caching (5-min, 1-hour) + auto caching (cache_control: {"type": "ephemeral"}). Memory tool (GA, API-only — not Claude.ai/Desktop).
Structured outputs (GA, incl. Bedrock): output_config.format or strict: true on tools. Schema limits: 20 strict tools, 24 optional params. output_format is deprecated.
Files API (beta): 500MB/file, 100GB/org. Citations: search_result blocks for RAG. Agent Skills (beta): doc generation (pptx, xlsx, docx, pdf) via code execution.
Tools: Bash, Text Editor, Computer Use, Web Search, Web Fetch, Code Execution, Memory, Tool Search (GA), MCP Connector (beta). Programmatic tool calling (GA). Dynamic filtering web tools (web_search_20260209, web_fetch_20260209). Code Execution free with web search/fetch.
Read references/api-features.md for configuration and code examples. Read references/tool-types.md for tool configs and compatibility matrix.
Architecture Decisions
When recommending architectures, apply these rules:
- Subagents vs tool chaining: Use subagents when each step needs its own system prompt or tool set. Use tool chaining when steps share context (under 200K tokens).
- MCP vs direct API: Use MCP when Claude needs runtime service access. Use direct API when you control orchestration.
- Batch vs streaming: Batch API for async workloads. Streaming for real-time UX. Programmatic tool calling to reduce round-trips.
- Model tiering: Haiku for high-volume classification, Sonnet for balanced tasks, Opus for complex reasoning. Add
effort: "low"for simple,"high"for critical steps. - Document processing: With 1M context, most documents fit in a single pass (prefer under 800K tokens). Only chunk when exceeding 1M or processing many docs in parallel (batch API). Multi-doc cross-referencing: load all if they fit, else subagents (one per doc -> structured summaries -> synthesis).
- Vision + Structured Outputs: Image content blocks +
output_config.format-> guaranteed JSON. - Batch + Prompt Caching: Combine for maximum savings. Cache persists across batch requests.
- 1M Context + Files API: Upload via Files API for large document processing.
- Memory + Compaction: Memory tool for critical facts + compaction API for long sessions.
Agent Capabilities
Agent SDK (Python, TypeScript — Claude Code and programmatic use only): query() for one-off tasks (supports hooks and custom tools), ClaudeSDKClient for persistent conversations, custom Transport for remote connections.
Subagents (Claude Code only): Agent tool (renamed from Task in v2.1.63). SendMessage auto-resumes stopped agents.
Hooks (Claude Code only): four types — command (shell), HTTP, prompt (LLM evaluation), and agent hooks. Events: Stop, StopFailure (API errors), PostCompact, Elicitation/ElicitationResult, InstructionsLoaded, ConfigChange, SessionStart.
Plugins (Claude Code only): bundle skills + agents + hooks + MCP + LSP servers + settings. ${CLAUDE_PLUGIN_DATA} for persistent state.
MCP Apps (beta): interactive HTML UIs. MCP elicitation: servers request structured input mid-task.
Read references/agent-capabilities.md for SDK API, custom tools, Agent Skills, and MCP Apps. Read references/claude-code-specifics.md for hooks, subagents, MCP config, and plugins.
Breaking Changes
When a user references any of these, proactively tell them the replacement:
| Deprecated | Replacement | Notes |
|---|---|---|
budget_tokens (Opus/Sonnet 4.6) |
thinking: {type: "adaptive"} + effort |
Still works on legacy models |
output_format |
output_config.format |
|
| Prefill (Opus 4.6) | Structured outputs | Returns 400 error |
| Sonnet 3.7, Haiku 3.5 | Sonnet 4.6, Haiku 4.5 | Retired Feb 2026 |
| Haiku 3 | Haiku 4.5 | Retiring Apr 2026 |
| Tool search beta header | Remove header | GA, no header needed |
| Code execution beta header | Remove header | GA, no header needed |
| Web search/fetch beta headers | Remove header | GA, no header needed |
| Memory tool beta header | Remove header | GA, no header needed |
Quick Reference (Key Parameters)
Adaptive thinking: thinking: {"type": "adaptive"} (Opus 4.6, Sonnet 4.6)
Thinking display omit: thinking: {"type": "adaptive", "display": "omitted"} (reduces response size, preserves signatures)
Extended thinking: thinking: {"type": "enabled", "budget_tokens": N} (Haiku 4.5, legacy)
Structured outputs: output_config: {"format": {"type": "json_schema", "schema": {...}}}
Effort: effort: "low" | "medium" | "high" | "max" (API max for Opus 4.6; Claude Code also supports max and auto)
Temperature: temperature: 0.0-1.0 (default 1.0)
Max tokens: max_tokens: N (required)
Streaming: stream: true
Memory tool: tools: [{"type": "memory_20250818", "name": "memory"}]
Web search: tools: [{"type": "web_search_20250305", "name": "web_search"}]
Web search + filtering: tools: [{"type": "web_search_20260209", ...}] (Opus/Sonnet 4.6)
Code execution: tools: [{"type": "code_execution_20250825", "name": "code_execution"}] (or code_execution_20260120 for programmatic calling)
Image input: {"type": "image", "source": {"type": "base64", "media_type": "image/png", "data": "..."}}
PDF input: {"type": "document", "source": {"type": "base64", "media_type": "application/pdf", "data": "..."}}
1M context: Native on Opus/Sonnet 4.6. Beta betas: ["context-1m-2025-08-07"] for Sonnet 4.5/4 (tier 4+)
Fast mode: speed: "fast", betas: ["fast-mode-2026-02-01"] (Opus 4.6, higher cost)
Auto caching: cache_control: {"type": "ephemeral"}
Files API: betas: ["files-api-2025-04-14"]
MCP connector: betas: ["mcp-client-2025-11-20"]
Reference Files
Read these when the user needs code examples or detailed configuration:
references/api-features.md— API config, beta headers, code for structured outputs, Files API, memory, citations, caching, streaming, fast modereferences/tool-types.md— Tool configurations, compatibility matrixreferences/agent-capabilities.md— Agent SDK API, custom tools, Agent Skills, MCP Appsreferences/model-specifics.md— Model IDs, per-model capabilities, provider details, migration guidesreferences/claude-code-specifics.md— Code Review, Remote Control, web sessions, Slack, background tasks, /loop, agent teams, CLI reference, IDE extensions, hooks, plugins, subagents, MCP config