┌─────────────────────────────────────────────────────────────┐
│                    COGNITIVE CONTROL                        │
├─────────────────────────────────────────────────────────────┤
│  IDENTITY      │ Who am I? What's my purpose?              │
│  CAPABILITIES  │ What tools can I use? What can't I do?    │
│  BEHAVIOR      │ How do I respond? When do I confirm?      │
│  GROUNDING     │ What prevents me from hallucinating?      │
│  ERRORS        │ How do I handle failures gracefully?      │
└─────────────────────────────────────────────────────────────┘

┌─────────────────────────────────────────┐
│ 1. IDENTITY & PURPOSE                   │
│    Who you are, what you do             │
├─────────────────────────────────────────┤
│ 2. CAPABILITIES & TOOLS                 │
│    What you CAN do (explicit list)      │
├─────────────────────────────────────────┤
│ 3. BOUNDARIES & CONSTRAINTS             │
│    What you CANNOT do (explicit list)   │
├─────────────────────────────────────────┤
│ 4. INTENT → TOOL MAPPING                │
│    How to interpret user requests       │
├─────────────────────────────────────────┤
│ 5. CONFIRMATION POLICY                  │
│    When to ask before acting            │
├─────────────────────────────────────────┤
│ 6. ERROR HANDLING                       │
│    How to handle failures gracefully    │
└─────────────────────────────────────────┘

## Intent Mapping

| User Says (Intent) | Tool to Call | Parameters | Confirmation |
|--------------------|--------------|------------|--------------|
| "create a task" | create_task | title, description | No |
| "delete task X" | delete_task | task_id | Yes |
| "show my tasks" | list_tasks | user_id | No |
| "mark X as done" | update_task | task_id, completed=true | No |

Before I [action], let me confirm:
- [Detail 1]
- [Detail 2]

Proceed? (yes/no)

## Grounding Rules

1. ONLY call tools listed in your capabilities
2. NEVER invent tool names or parameters
3. If no tool matches the request, say so clearly
4. NEVER claim to have done something you didn't do
5. If a tool call fails, report the failure honestly

When you receive a request you cannot fulfill:

1. Acknowledge the request
2. Explain what you CAN do instead
3. Offer alternatives if available
4. Never pretend or hallucinate capabilities

## When Tools Fail

If a tool returns an error:
1. Do NOT retry automatically (unless specified)
2. Explain what happened in plain language
3. Suggest what the user can do
4. Offer to try an alternative approach

Example:
"I tried to create the task, but the database returned an error.
This might be a temporary issue. Would you like me to try again,
or would you prefer to wait a few minutes?"

## Graceful Degradation

If you cannot complete a request:
1. Complete as much as possible
2. Clearly state what succeeded and what failed
3. Provide partial results if available
4. Suggest next steps

1. CLARIFY
   └─→ Gather identity, tools, safety requirements

2. ENUMERATE TOOLS
   └─→ List all available tools with descriptions

3. MAP INTENTS
   └─→ Create intent → tool mapping table

4. DEFINE BOUNDARIES
   └─→ List forbidden actions and scope limits

5. SET CONFIRMATION POLICY
   └─→ Mark which actions need confirmation

6. ADD ERROR HANDLING
   └─→ Define failure responses

7. ASSEMBLE PROMPT
   └─→ Use template to create complete prompt

8. VALIDATE
   └─→ Check for gaps, ambiguities, conflicts