diff --git a/dynamiq/nodes/agents/agent.py b/dynamiq/nodes/agents/agent.py
index 8c646bbaa..0f61cd045 100644
--- a/dynamiq/nodes/agents/agent.py
+++ b/dynamiq/nodes/agents/agent.py
@@ -1576,6 +1576,14 @@ def _init_prompt_blocks(self):
             has_sub_agent_tools=any(isinstance(t, SubAgentTool) for t in self.tools),
         )
 
+        # Append user-provided instructions to operational_instructions block
+        if self.instructions:
+            existing = self.system_prompt_manager._prompt_blocks.get("operational_instructions", "")
+            self.system_prompt_manager.set_block(
+                "operational_instructions",
+                f"{existing}\n\n{self.instructions}" if existing else self.instructions,
+            )
+
         # Only auto-wrap the entire role in a raw block if the user did not
         # provide explicit raw/endraw markers. This allows roles to mix
         # literal sections (via raw) with Jinja variables like {{ input }}
diff --git a/dynamiq/nodes/agents/base.py b/dynamiq/nodes/agents/base.py
index d20bb9b8e..ae20e1c64 100644
--- a/dynamiq/nodes/agents/base.py
+++ b/dynamiq/nodes/agents/base.py
@@ -236,6 +236,10 @@ class Agent(Node):
             Can be used to provide additional context or instructions to the agent.
             Accepts Jinja templates to provide additional parameters.""",
     )
+    instructions: str | None = Field(
+        default=None,
+        description="Additional operational instructions appended to the operational instructions block.",
+    )
     description: str | None = Field(default=None, description="Short human-readable description of the agent.")
     _mcp_servers: list[MCPServer] = PrivateAttr(default_factory=list)
     _excluded_tool_ids: set[str] = PrivateAttr(default_factory=set)
diff --git a/dynamiq/nodes/agents/prompts/manager.py b/dynamiq/nodes/agents/prompts/manager.py
index 4db84eb47..1f4a59279 100644
--- a/dynamiq/nodes/agents/prompts/manager.py
+++ b/dynamiq/nodes/agents/prompts/manager.py
@@ -14,7 +14,7 @@
     REACT_BLOCK_INSTRUCTIONS_STRUCTURED_OUTPUT,
     REACT_BLOCK_OUTPUT_FORMAT,
     REACT_BLOCK_TOOLS,
-    REACT_BLOCK_TOOLS_NO_FORMATS,
+    REACT_BLOCK_TOOLS_BRIEF,
     REACT_BLOCK_XML_INSTRUCTIONS_NO_TOOLS,
     REACT_BLOCK_XML_INSTRUCTIONS_SINGLE,
     REACT_MAX_LOOPS_PROMPT,
@@ -321,7 +321,7 @@ def get_model_specific_prompts(
             )
             if has_tools:
                 prompt_blocks["tools"] = get_prompt_constant(
-                    model_name, "REACT_BLOCK_TOOLS_NO_FORMATS", REACT_BLOCK_TOOLS_NO_FORMATS
+                    model_name, "REACT_BLOCK_TOOLS_BRIEF", REACT_BLOCK_TOOLS_BRIEF
                 )
 
         case InferenceMode.STRUCTURED_OUTPUT:
@@ -341,7 +341,7 @@ def get_model_specific_prompts(
                 logger.debug(f"Using model-specific REACT_BLOCK_XML_INSTRUCTIONS_SINGLE for '{model_name}'")
             prompt_blocks["instructions"] = xml_instructions_no_tools if not has_tools else instructions_xml
 
-    # Build secondary_instructions from enabled features
+    # Build operational_instructions from enabled features
     secondary_parts = []
     if parallel_tool_calls_enabled:
         secondary_parts.append(REACT_BLOCK_MULTI_TOOL_PLANNING)
@@ -354,16 +354,15 @@ def get_model_specific_prompts(
         secondary_parts.append(CONTEXT_MANAGER_INSTRUCTIONS)
     if todo_management_enabled:
         secondary_parts.append(TODO_TOOLS_INSTRUCTIONS)
-    if sandbox_base_path:
-        secondary_parts.append(
-            SANDBOX_INSTRUCTIONS_TEMPLATE.format(
-                base_path=sandbox_base_path,
-            )
-        )
     if has_sub_agent_tools:
         secondary_parts.append(SUB_AGENT_INSTRUCTIONS)
 
     if secondary_parts:
-        prompt_blocks["secondary_instructions"] = "\n\n".join(secondary_parts)
+        prompt_blocks["operational_instructions"] = "\n\n".join(secondary_parts)
+
+    if sandbox_base_path:
+        prompt_blocks["environment"] = SANDBOX_INSTRUCTIONS_TEMPLATE.format(
+            base_path=sandbox_base_path,
+        )
 
     return prompt_blocks, agent_template
diff --git a/dynamiq/nodes/agents/prompts/react/instructions.py b/dynamiq/nodes/agents/prompts/react/instructions.py
index 905f6d340..db162cf2b 100644
--- a/dynamiq/nodes/agents/prompts/react/instructions.py
+++ b/dynamiq/nodes/agents/prompts/react/instructions.py
@@ -33,12 +33,12 @@
 - Explicitly link key statements to specific findings from the referenced materials to strengthen credibility and transparency.
 - Make sure to adhere to AGENT PERSONA & STYLE & ADDITIONAL BEHAVIORAL GUIDELINES.
 
-SINGLE ACTION PER TURN:
+## Single Action Per Turn
 - Execute exactly ONE action per response, then wait for its Observation before continuing
 - Do NOT chain multiple Action/Action Input pairs in the same response
 - After receiving an Observation, decide the next single action based on the result
 
-FILE HANDLING:
+## File Handling
 - Tools may generate or process files (images, CSVs, PDFs, etc.)
 - If you want to return files, include an "Output Files:" line before "Answer:" listing file paths (comma-separated). This line is optional — omit it if there are no files to return.
 """  # noqa: E501
@@ -82,7 +82,7 @@
     <output_files>[Optional: comma-separated absolute file paths to return]</output_files>
 </output>
 
-CRITICAL XML FORMAT RULES:
+## Critical XML Format Rules
 - ALWAYS include <thought> tags with detailed reasoning
 - Start the text immediately after each opening tag; do not add leading newlines or indentation inside the tags
 - Write thoughts in the first person (e.g., "I will...", "I should...")
@@ -105,19 +105,19 @@
 - Explicitly link key statements to specific findings from the referenced materials to strengthen credibility and transparency.
 - Make sure to adhere to AGENT PERSONA & STYLE & ADDITIONAL BEHAVIORAL GUIDELINES.
 
-SINGLE ACTION PER TURN:
+## Single Action Per Turn
 - Execute exactly ONE <action>/<action_input> pair per response, then wait for its Observation before continuing
 - Do NOT include multiple action blocks or answer blocks in the same response
 - After receiving an Observation, decide the next single action based on the result
 
-JSON FORMATTING REQUIREMENTS:
+## JSON Formatting Requirements
 - Put JSON on single line within tags
 - Use double quotes for all strings
 - Escape newlines as \\n, quotes as \\"
 - NO multi-line JSON formatting
 - For tools that accept code (e.g. python), the code must be one JSON string with \\n for line breaks, not literal newlines
 
-FILE HANDLING:
+## File Handling
 - Tools may generate or process files (images, CSVs, PDFs, reports, etc.)
 - If you want to return files, include an <output_files> tag after </answer> (but still inside <output>) listing absolute file paths (comma-separated). This tag is optional — omit it if there are no files to return.
 """  # noqa: E501
@@ -145,6 +145,11 @@
 {{ tool_description }}
 """
 
+REACT_BLOCK_TOOLS_BRIEF = """
+Available tools: [{{ tools_name }}]
+Refer to each tool's function schema for detailed usage.
+"""
+
 
 REACT_BLOCK_INSTRUCTIONS_STRUCTURED_OUTPUT = """Always structure your responses in this JSON format:
 
@@ -179,7 +184,7 @@
 - Ensure proper JSON syntax with quoted keys and values
 - To return an agent tool's response as the final output, include "delegate_final": true inside that tool's action_input. Use this only for a single agent tool call and do not call finish yourself afterward; the system will return the agent's result directly.
 
-FILE HANDLING:
+## File Handling
 - Tools may generate or process files (images, CSVs, PDFs, etc.)
 - When using action "finish", include an "output_files" field with comma-separated file paths to return. Use an empty string if there are no files to return.
 - Never return empty response.
@@ -197,7 +202,7 @@
 Only after utilizing the necessary tools and gathering the required information should
 you call `provide_final_answer` to deliver the final response.
 
-FUNCTION CALLING GUIDELINES:
+## Function Calling Guidelines
 - ALWAYS populate the "thought" field FIRST before "action_input" in your function calls
 - Analyze the request carefully to determine if tools are needed
 - Call functions with properly formatted arguments
@@ -205,7 +210,7 @@
 - Chain multiple tool calls when necessary for complex tasks
 - If you want an agent tool's response returned verbatim as the final output, include "delegate_final": true inside that tool's action_input. Use this only for a single agent tool call and do not call provide_final_answer yourself; the system will return the agent's result directly.
 
-FILE HANDLING:
+## File Handling
 - Tools may generate or process files (images, CSVs, PDFs, etc.)
 - When calling `provide_final_answer`, include an `output_files` argument with comma-separated file paths to return. Pass an empty string if there are no files to return.
 """  # noqa: E501
diff --git a/dynamiq/nodes/agents/prompts/secondary_instructions.py b/dynamiq/nodes/agents/prompts/secondary_instructions.py
index 4f5d2cd05..e4ed353cf 100644
--- a/dynamiq/nodes/agents/prompts/secondary_instructions.py
+++ b/dynamiq/nodes/agents/prompts/secondary_instructions.py
@@ -1,23 +1,26 @@
 """Secondary instructions for agent prompts."""
 
 DELEGATION_INSTRUCTIONS = (
-    "- Optional: If you want an agent tool's response returned verbatim as the final output, "
-    'set "delegate_final": true in that tool\'s input. Use this only for a single agent tool call '
-    "and do not provide your own final answer; the system will return the agent's result directly. "
-    "Do not set delegate_final: true inside metadata of the input, it has to be a separate field."
+    "## Delegation\n"
+    "- If you want an agent tool's response returned verbatim as the final output, "
+    'set "delegate_final": true in that tool\'s input.\n'
+    "- Use this only for a single agent tool call "
+    "and do not provide your own final answer; the system will return the agent's result directly.\n"
+    "- Do not set delegate_final: true inside metadata of the input, it has to be a separate field."
 )
 
 DELEGATION_INSTRUCTIONS_XML = (
+    "## Delegation\n"
     '- To return an agent tool\'s response as the final output, include "delegate_final": true inside that '
     "tool's <input> or <action_input>. Use this only for a single agent tool call and do not provide an "
     "<answer> yourself; the system will return the agent's result directly."
 )
 
-CONTEXT_MANAGER_INSTRUCTIONS = """CONTEXT MANAGEMENT:
+CONTEXT_MANAGER_INSTRUCTIONS = """## Context Management
 - Use the context-manager tool proactively when conversation is getting long
 - Save critical info (IDs, filenames) in "notes" field BEFORE calling - previous messages will be summarized"""
 
-TODO_TOOLS_INSTRUCTIONS = """TODO MANAGEMENT:
+TODO_TOOLS_INSTRUCTIONS = """## Todo Management
 - Use the todo-write tool for complex 3+ step tasks; skip for simple requests
 - Current todos shown in [State: ...] at the end of user last messages under "Todos:"
 - When creating initial list: first task "in_progress", rest "pending"
@@ -26,7 +29,7 @@
 - Only mark completed when FULLY done; if blocked, keep in_progress"""
 
 
-SANDBOX_INSTRUCTIONS_TEMPLATE = """SANDBOX EXECUTION ENVIRONMENT:
+SANDBOX_INSTRUCTIONS_TEMPLATE = """## Sandbox Environment
 You operate inside a persistent sandbox filesystem.
 The sandbox directory is your working memory.
 
@@ -35,7 +38,7 @@
 - Uploaded files are placed in {base_path}/input/.
 - Other tools can ONLY access files under {base_path}/.
 
-CRITICAL PRINCIPLE — PERSIST EVERYTHING IMPORTANT:
+## Persistence
 The sandbox filesystem is your long-term memory. Do not rely on conversation context alone.
 
 Whenever you:
@@ -53,7 +56,7 @@
 - parsed_results.md
 - intermediate_analysis_step_1.md
 
-EXECUTION RULES:
+## Execution Rules
 1. Use 'python3' instead of 'python'.
 2. For Python tasks:
    - Always write a .py script file first; then execute it.
@@ -74,123 +77,23 @@
 """
 
 
-REACT_BLOCK_MULTI_TOOL_PLANNING = """
-MULTI-TOOL PLANNING AND STRATEGY:
-
-Core Principle: Scale tool usage to match task complexity
-Start with the minimum number of tools needed and scale up based on the task's requirements.
-
-Decision Framework:
-1. Zero Tools - Answer directly when:
-   - Question is within your knowledge base
-   - No real-time data needed
-   - Simple explanations or general concepts
-
-2. Single Tool - Use one tool when:
-   - Single fact verification needed
-   - One specific data point required
-   - Simple lookup or calculation
-
-3. Multiple Tools (2-4) - Use parallel tools when:
-   - Comparing information from different sources
-   - Gathering complementary data points
-   - Cross-referencing or validation needed
-
-4. Comprehensive Research (5+) - Use extensive tooling when:
-   - Deep analysis requested ("comprehensive", "detailed", "thorough")
-   - Multiple aspects of complex topic
-   - Creating reports or extensive documentation
-
-MANDATORY MULTI-TOOL PATTERNS:
-
-1. Research & Data Tasks:
-
-   - Start: Analyze query complexity.
-   - Simple facts / single source: 1–2 tool calls.
-   - Moderate complexity / multiple sources: 3–4 tool calls with complementary queries.
-   - Comprehensive research / broad topics: 5+ tool calls covering:
-     - General overview
-     - Specific benefits, use cases, or features
-     - Limitations, challenges, or drawbacks
-     - Comparisons or alternatives
-     - Recent developments or updates
-   - Example progression:
-     1. General topic overview
-     2. Specific aspects (benefits, use cases)
-     3. Challenges or limitations
-     4+. Deep dives on critical aspects
-
-2. Coding & Technical Tasks:
-   - Documentation lookup: 1–2 tools (official docs + examples)
-   - Debugging: 2–3 tools (error search + solution patterns)
-   - Architecture decisions: 3–5 tools (best practices + comparisons + examples)
-   - Full implementation: 5+ tools (docs + patterns + edge cases + optimization)
-
-3. Comparative Analysis / Verification:
-   - Single source: 1 tool
-   - Multiple sources: Use parallel calls for efficiency
-   - Comparative or critical analysis: Minimum 3 sources
-   - Market research or controversial topics: 3+ diverse sources, cross-referenced
-   - Scientific/technical validation: Cross-check experiments, methods, or datasets
-
-EFFICIENCY GUIDELINES:
-
-1. Parallel vs Sequential Tool Calls:
-   - Use PARALLEL calls when queries are independent
-   - Use SEQUENTIAL only when later queries depend on earlier results
-
-2. Optimized Query Strategy:
-   - Start broad, then narrow based on results.
-   - Use varied search parameters to cover different angles.
-   - Each tool call should add unique value; avoid redundant queries.
-   - Scale detail dynamically: begin with essentials, add deeper queries if needed,
-   and stop once sufficient information is gathered.
-
-MULTIPLE ENTITIES PATTERN:
-When researching multiple distinct entities (companies, products, people, locations):
-- BAD: "Report about Company A, Company B, Company C"
-- GOOD: Separate queries for each entity
-  - Query 1: "Company A financial performance"
-  - Query 2: "Company B market position"
-  - Query 3: "Company C recent developments"
-- REASON: Each entity gets full search attention, avoiding diluted results
-
-QUERY FORMULATION BEST PRACTICES:
-1. Entity Separation: Search distinct subjects individually
-2. Perspective Variation: Use different angles for same topic
-   - "benefits of X" → "advantages of X" → "X success stories"
-3. Temporal Layering: Mix timeframes
-   - "latest developments in X"
-   - "X trends 2025-2026"
-   - "future of X"
-4. Source Diversification: Target different source types
-   - Official documentation
-   - Independent reviews/analysis
-   - Case studies/examples
-
-REFLECTION AND VALIDATION:
-
-- Before and after each tool use, briefly reflect on whether it's the most appropriate choice, how it contributes to solving the problem,
-and whether the results are relevant, complete, and high quality – adjusting your strategy if needed.
-- Before delivering the final answer, confirm that all parts of the original question are addressed
-and that your conclusions are well-supported and internally consistent.
-
-IMPORTANT RULES:
-- Quality over quantity - each tool call must serve a purpose
-- Explain your multi-tool strategy in your thought process
-- Adapt the number of tools based on result quality
-- If initial results are comprehensive, don't add unnecessary calls
-- For coding: balance between documentation, examples, and best practices
-- Always consider user's implicit needs beyond explicit request
-- Employ strategic thinking and reflection at each step
-"""  # noqa: E501
-
-SUB_AGENT_INSTRUCTIONS = """SUB-AGENT TOOLS:
+REACT_BLOCK_MULTI_TOOL_PLANNING = """## Multi-Tool Planning
+
+- Scale tool usage to task complexity: answer directly when possible, use 1-2 tools for simple lookups,
+  3-5 for multi-source research, more for comprehensive analysis.
+- Use PARALLEL calls for independent queries; SEQUENTIAL only when results feed into the next call.
+- When researching multiple entities, search each one separately - don't combine into one query.
+- Each tool call must add unique value. Avoid redundant or semantically similar queries.
+- Start broad, then narrow. Stop once sufficient information is gathered.
+- Verify that all parts of the original question are addressed before delivering the final answer.
+"""
+
+SUB_AGENT_INSTRUCTIONS = """## Sub-Agent Tools
 Sub-agent tools are specialized agents you can delegate tasks to. Each sub-agent is an independent agent \
-with its own tools and expertise — use them to break complex work into focused subtasks.
+with its own tools and expertise - use them to break complex work into focused subtasks.
 
 Execution modes (shown in each tool's description):
-- "[Independent agent: ...]" — spawns a fresh instance per call. Safe to call in parallel with other tools.
-- "[Shared agent: ...]" — reuses a single instance. Calls run sequentially; do not call in parallel.
+- "[Independent agent: ...]" - spawns a fresh instance per call. Safe to call in parallel with other tools.
+- "[Shared agent: ...]" - reuses a single instance. Calls run sequentially; do not call in parallel.
 
 Provide each sub-agent with a clear, self-contained task description in the "input" field."""
diff --git a/dynamiq/nodes/agents/prompts/templates.py b/dynamiq/nodes/agents/prompts/templates.py
index 7f9168808..12e5ec2cd 100644
--- a/dynamiq/nodes/agents/prompts/templates.py
+++ b/dynamiq/nodes/agents/prompts/templates.py
@@ -1,78 +1,88 @@
 AGENT_PROMPT_TEMPLATE = """
 You are AI powered assistant.
 
+## Behaviour
+- Be concise and direct. No unnecessary preamble or narration - just act.
+- If the request is underspecified, ask only the minimum needed to proceed.
+- Focus on what’s true, not on agreeing.
+- If something is wrong, point it out clearly and respectfully.
+
+## Doing Tasks
+1. Think - break down what is needed before acting.
+2. Execute - use tools to deliver the result.
+3. Validate - check that the output fully satisfies the request.
+Do not stop until the task is complete.
+
+
 {%- if instructions %}
-# PRIMARY INSTRUCTIONS
+---
+[1] PRIMARY INSTRUCTIONS
 {{instructions}}
-
 {%- endif %}
 
-{%- if secondary_instructions %}
-# SECONDARY INSTRUCTIONS
-{{secondary_instructions}}
+{%- if environment %}
+---
+[2] ENVIRONMENT
+{{environment}}
+{%- endif %}
 
+{%- if operational_instructions %}
+---
+[3] OPERATIONAL INSTRUCTIONS
+{{operational_instructions}}
 {%- endif %}
 
 {%- if tools %}
-# AVAILABLE TOOLS
+---
+[4] AVAILABLE TOOLS
 {{tools}}
-
 {%- endif %}
 
 {%- if skills %}
-# AVAILABLE SKILLS (skills-tool)
+---
+[5] AVAILABLE SKILLS (skills-tool)
 The list below is for orientation only.
-You must read full skill content before applying it;
-do not rely on the short descriptions.
+Read full skill content before applying it; do not rely on short descriptions.
 
 {{skills}}
 
-## How to read skill content
+How to read skill content:
 {%- if sandbox_skills_base_path %}
-- The list above includes the path to each skill in the sandbox.
-  Read content via sandbox-shell.
+- The list above includes the path to each skill in the sandbox. Read content via sandbox-shell.
 - Keep reads targeted to avoid large content:
-  Prefer grep and line ranges over dumping the whole file.
-  (1) Use `grep -n "## Section title" <path>` to find line numbers of sections,
-      then read only that range with `sed -n 'START,ENDp' <path>`.
-  (2) Or search for a keyword with `grep -n "keyword" <path>`
-      and read a window around it.
-  (3) For a quick overview, use `head -n 150 <path>`
-      then read more sections as needed.
-  Only use `cat <path>` when the skill is short or
-  you have confirmed you need the full content.
-- Scripts are under {{ sandbox_skills_base_path }}/<skill_name>/scripts/ — run them via the sandbox.
+  1. Use `grep -n "## Section title" <path>` to find line numbers, then `sed -n 'START,ENDp' <path>`.
+  2. Or search for a keyword with `grep -n "keyword" <path>` and read a window around it.
+  3. For a quick overview, use `head -n 150 <path>` then read more as needed.
+  Only use `cat <path>` when the skill is short or you need the full content.
+- Scripts are under {{ sandbox_skills_base_path }}/<skill_name>/scripts/ - run them via the sandbox.
 {%- else %}
-- Use skills-tool: Call action="list" to see available skills, then action="get" with skill_name="..."
-  to load the full skill instructions.
-  For large skills use section="Section title" or line_start/line_end (1-based) to read only a part.
+- Use skills-tool: action="list" to see available skills,
+  then action="get" with skill_name="..." to load full instructions.
+  For large skills use section="Section title" or line_start/line_end
+  (1-based) to read only a part.
 {%- endif %}
-- After reading (from sandbox or get), apply the skill's instructions yourself
-  in your reasoning and provide the result in your final answer.
+- After reading, apply the skill's instructions in your reasoning and provide the result in your final answer.
   Do not call the tool again with user content to transform.
-
 {%- endif %}
 
 {%- if output_format %}
-# RESPONSE FORMAT
+---
+[6] RESPONSE FORMAT
 {{output_format}}
-
 {%- endif %}
 
 {%- if role %}
-# AGENT PERSONA & STYLE & ADDITIONAL BEHAVIORAL GUIDELINES
-(This section defines how the assistant presents information and behaves - its personality, tone, communication style,
-and additional behavioral guidelines.
-These supplementary instructions enhance the agent's interactions but should never override or contradict
-the PRIMARY INSTRUCTIONS above.)
+---
+[7] AGENT PERSONA
+This section defines persona, tone, and behavioral guidelines.
+These MUST NOT override or contradict [1] PRIMARY INSTRUCTIONS.
 {{role}}
-
 {%- endif %}
 
 {%- if context %}
-# CONTEXT
+---
+[8] CONTEXT
 {{context}}
-
 {%- endif %}
 
 {%- if date %}