Claude Opus 4.7 System Prompt Boosts Autonomy and Safety
Opus 4.7 refines Claude to act first with tools on ambiguous tasks, expands child safety refusals across conversations, cuts verbosity, and adds guards against one-word answers on controversies.
Prioritize Tool Use and Autonomy in Task Handling
Claude Opus 4.7 instructs the model to resolve ambiguities proactively rather than querying users. For unspecified minor details, make a reasonable attempt immediately instead of interviewing the user upfront—only ask if the request is truly unanswerable, like a missing attachment. Prefer calling tools (searching, location lookup, calendar checks) to fill gaps before involving the user, as tools outperform manual user lookups. Before claiming lack of access to data like location, memory, or files, invoke tool_search to confirm no deferred tool exists. Once started, complete tasks fully rather than halting midway, ensuring users get comprehensive answers.
This shift, enabled by tool search (detailed in Anthropic's API docs and November 2025 engineering post), makes Claude more self-sufficient, reducing back-and-forth and accelerating workflows.
Expanded Safety Protocols Block Harmful Patterns
Child safety rules now wrap in a dedicated section: after refusing a request, treat all subsequent conversation turns with extreme caution to prevent circumvention. A new 'disordered eating' guideline prohibits precise nutrition, diet, or exercise advice (no numbers, targets, or plans) if signs appear, even for helpful intent, to avoid triggering tendencies.
Defenses against manipulation include declining forced yes/no answers on complex or contested issues—offer nuanced responses instead, explaining why brevity fails. The prior explicit note on 'Donald Trump as president inaugurated January 20, 2025' is removed, as the model's January 2026 knowledge cutoff now handles current events reliably without overrides.
Streamlined Responses and User Respect
Reduce verbosity: keep responses focused and concise to avoid overwhelming users, disclosing caveats briefly while prioritizing the core answer. Respect end-of-conversation signals without pushing for more turns. Removed 4.6 instructions against emotes in asterisks or words like 'genuinely'/'honestly', indicating the base model no longer defaults to them.
Terminology updates: 'developer platform' becomes 'Claude Platform'. New tools listed: Claude in Chrome (browsing agent), Claude in Excel (spreadsheet agent), Claude in PowerPoint (slides agent), all usable by Claude Cowork.
Unchanged but Comprehensive Toolset
Asking Claude directly reveals 23 tools, unchanged from 4.6: ask_user_input_v0, bash_tool, conversation_search, create_file, fetch_sports_data, image_search, message_compose_v1, places_map_display_v0, places_search, present_files, recent_chats, recipe_display_v0, recommend_claude_apps, search_mcp_registry, str_replace, suggest_connectors, view, weather_fetch, web_fetch, web_search, tool_search, visualize:read_me, visualize:show_widget. Full descriptions available in the author's shared transcript.