LLM Wiki — Synaptic Lattice

Anthropic Research Instructions

Properties1
tagsprompt-engineering, system-prompt, tool-origin, research, workflow-pattern

Leitura de 4 min

Anthropic Research Instructions

This capture presents the leaked system prompt for Anthropic’s advanced research mode — a runtime configuration applied to Claude when a user enables “advanced research” in their interface settings. Unlike the consumer chat prompt ([[claude-opus-5-system-prompt]]) or the CLI coding agent prompt ([[claude-code-system-prompt]]), this prompt is injected as a behavioral overlay that re-routes nearly all queries through a single tool: launch_extended_search_task.

The source is confidence: low because it is a leaked-prompt artifact from a public mirror (asgeirtj/system_prompts_leaks), not an authenticated vendor publication. It is useful for comparative pattern analysis of Anthropic’s research-mode prompt design.

Tool-Priority Override

The prompt opens with a <tool_selection_instructions> block containing a CRITICAL INSTRUCTION: launch_extended_search_task MUST be used by default for everything except (1) basic conversational messages and (2) extremely simple factual questions. This is a rare “tool monopoly” design — when the tool is present, it takes absolute priority over ALL other tools, including web_search, file-reading tools, and any other available function.

The rationale is encoded in the prompt itself: the user “intentionally enabled this tool, so they clearly expect Claude to use it by default and will be upset if Claude does not use the research tool.” This is product-design-by-prompt rather than by API schema restriction.

Implications

A system prompt can enforce tool routing without backend gatekeeping. Authors of agent systems can adopt tool-priority overrides to steer behavior toward a specific capability surface — useful when a premium tool (research, code execution, database access) should dominate the default workflow.

Clarifying-Questions Protocol

Before launching research, Claude is permitted to ask up to 3 clarifying questions, governed by strict rules: (1) never ask for confirmation when the query is already clear and specific; (2) pick reasonable defaults and note assumptions rather than asking when details are unspecified; (3) only ask when the answer would send research in a completely different direction. After one round of questions, Claude MUST immediately launch — never ask twice.

The prompt includes worked examples (revenue-growth startup query, Protestant-pastor shortage query, email-summary query, Taiwan-travel query) showing the threshold for when questions are warranted versus when to launch immediately.

Implications

The protocol turns conversational friction into a structured gate. The “never ask twice” rule prevents clarification loops — a pattern worth porting to any agent that asks questions before acting.

The <mandatory_copyright_requirements> block forbids reproducing any copyrighted material from search results (max one quote under 20 words per result), never reproducing song lyrics, and flagging copyright-sensitive requests before researching. The <harmful_content_safety> block forbids using or citing sources that promote hate, violence, extremism, self-harm, misinformation, unauthorized pharmaceuticals, or surveillance — even if the user requests them.

Implications

Safety and copyright compliance are embedded directly in the system prompt as hard constraints on the research workflow, not post-hoc filters. This makes the prompt itself a compliance layer.

Envelope and Tool Schema

The prompt uses an XML-tag envelope (<research_instructions>, <tool_selection_instructions>, <clarifying_questions_rules>, <good_examples>, <search_response_guidelines>, <mandatory_copyright_requirements>, <harmful_content_safety>, <critical_reminders>) wrapping markdown content. A trailing <function> block declares the launch_extended_search_task tool as JSON schema, with output_markdown_artifact and output_react_artifact booleans defaulting to false and triggered only by explicit user phrases (“subagent markdown artifact” / “react artifact”).

Implications

The XML-tag envelope keeps instruction domains visually separable for the model, while the embedded JSON schema makes the tool contract explicit — a hybrid the prompt wiki can classify as envelope_format: xml.

Open Questions

  • Whether the output_markdown_artifact / output_react_artifact flags are surfaced to end-users as visible options or only parseable by the model as magic phrases.
  • How the <search_response_guidelines> note about “user’s location (provided in user context below)” is populated at runtime — the raw prompt contains no location data, only a placeholder reference.
  • Whether the “tool monopoly” override is reversible within a session (can the user demand a different tool) or whether the prompt makes launch_extended_search_task truly inescapable.

Relationships

Sources

  • raw/prompts/leaked/anthropic-research-instructions.md

Visão de gráfico

Backlinks