ReddRadar
Agent skill
agent-prompt
Write effective system prompts for TD AI agents. Covers role definition, constraint specification, output formatting, tool usage instructions, and prompt structure patterns.
Install this agent skill to your Project
npx add-skill https://github.com/treasure-data/td-skills/tree/main/tdx-skills/agent-prompt
SKILL.md
Writing Effective Agent Prompts
Guidelines for crafting system prompts that produce reliable, consistent agent behavior.
Prompt Structure
# Role Definition
[Who the agent is and what it does]
# Constraints
[What the agent MUST and MUST NOT do]
# Process/Workflow
[Step-by-step instructions]
# Output Format
[Expected response structure]
# Context
[Environment info, available tools]
Key Principles
1. Define Role Clearly
You are a customer support specialist for [Company].
Your role is to help customers resolve billing issues
and answer questions about their accounts.
Not: "You are a helpful assistant."
2. Use Explicit Constraints
DO this:
=== CRITICAL CONSTRAINTS ===
You are STRICTLY PROHIBITED from:
- Modifying customer payment methods without verification
- Sharing account details with third parties
- Making promises about refunds over $100
You MUST:
- Verify customer identity before account changes
- Log all interactions in the CRM
NOT this:
Be careful with customer data.
3. Provide Step-by-Step Process
## Your Process
1. **Verify Identity**: Ask for account email and last 4 digits of phone
2. **Understand Issue**: Let customer explain before offering solutions
3. **Check History**: Review past interactions in CRM
4. **Propose Solution**: Offer specific resolution with timeline
5. **Confirm**: Summarize actions taken and next steps
4. Specify Output Format
## Required Output
End your response with:
### Summary
- Issue: [Brief description]
- Resolution: [Action taken]
- Follow-up: [Next steps if any]
Format all dates as YYYY-MM-DD.
Use bullet points for lists of 3+ items.
5. Handle Edge Cases
## Special Cases
If customer is angry:
- Acknowledge frustration first
- Do not argue or justify
- Escalate if they mention legal action
If request is outside scope:
- Politely explain limitations
- Provide alternative resources
- Offer to transfer to appropriate team
6. Tool Usage Instructions
## Available Tools
You have access to these tools:
- `search_kb`: Search knowledge base for product info
- `lookup_customer`: Get customer account details
- `create_ticket`: Create support ticket
ALWAYS use `lookup_customer` before discussing account details.
NEVER make up information - use `search_kb` when unsure.
Common Patterns
Read-Only Agent
=== READ-ONLY MODE ===
You can ONLY read and analyze. You CANNOT:
- Create, modify, or delete any records
- Execute transactions
- Change system state
Your role is EXCLUSIVELY to provide information and recommendations.
Multi-Step Workflow Agent
## Workflow
1. **Gather**: Collect all required information
2. **Validate**: Check inputs against rules
3. **Process**: Execute the main task
4. **Verify**: Confirm results are correct
5. **Report**: Summarize what was done
Do NOT skip steps. If any step fails, stop and report the issue.
Expert Agent
You are a [domain] expert with deep knowledge of:
- [Specific area 1]
- [Specific area 2]
- [Specific area 3]
When answering:
- Cite specific sources or documentation
- Provide concrete examples
- Acknowledge uncertainty when present
Anti-Patterns to Avoid
| Bad | Good |
|---|---|
| "Be helpful" | "Answer questions about X using the knowledge base" |
| "Don't be rude" | "Use professional, neutral tone" |
| "Handle errors appropriately" | "If error occurs, log it and notify user with error code" |
| "Use good judgment" | "If amount > $100, require manager approval" |
| Long paragraphs | Bulleted lists and numbered steps |
Variables in Prompts
Use {{variable}} for dynamic content:
You are assisting {{customer_name}} (ID: {{customer_id}}).
Their current plan is {{plan_type}}.
Focus on issues related to their plan tier.
Testing Prompts
Before deploying, test with:
- Happy path: Normal expected inputs
- Edge cases: Empty inputs, special characters
- Adversarial: Attempts to break constraints
- Out of scope: Requests outside agent's domain
prompt.md Example
You are a SQL query assistant for Treasure Data.
## Your Role
Help users write and optimize Trino SQL queries for TD.
## Constraints
- NEVER execute queries that modify data (INSERT, UPDATE, DELETE)
- ALWAYS include time filters using TD_INTERVAL or TD_TIME_RANGE
- Do NOT guess table schemas - use `describe_table` tool first
## Process
1. Understand the user's data question
2. Check available tables with `list_tables`
3. Get schema with `describe_table`
4. Write query with appropriate time filters
5. Explain the query logic
## Output Format
Provide queries in code blocks with explanation:
```sql
SELECT ...
Explanation: [Why this query answers the question]
## Related Skills
- **agent** - Agent configuration and deployment
- **tdx-basic** - CLI operations
Recommended Agent Skills
Expand your agent's capabilities with these related and highly-rated skills.
treasure-data/td-skills
email-campaign
This skill should be used when the user asks to "create an email", "build an email campaign", "design an email template", "generate an email for a segment", "preview an email", or "push an email to Engage". Generates enterprise-grade HTML email templates with live preview in Treasure Studio and natural language editing, then pushes the final version to Treasure Engage.
treasure-data/td-skills
action-report
YAML format reference for action reports rendered via preview_action_report. MUST be read before writing any action report YAML — defines the report structure (title, summary, actions array) and action item fields (as_is, to_be, reason, priority, category, impact) with incremental build workflow. Required by seo-analysis and any skill that produces prioritized recommendations.
treasure-data/td-skills
grid-dashboard
YAML format reference for grid dashboards rendered via preview_grid_dashboard. MUST be read before writing any dashboard YAML — defines the page structure, 6 cell types (kpi, gauge, scores, table, chart, markdown), grid layout rules, cell merging syntax, and incremental build workflow. Required by seo-analysis and any skill that produces visual data dashboards.
treasure-data/td-skills
seo-analysis
Runs SEO and AEO (Answer Engine Optimization) analysis on websites or specific pages. Use when the user mentions SEO, AEO, search rankings, search optimization, or wants to analyze how their pages perform in search engines and AI answers. Produces a data dashboard and action report with before/after recommendations.
treasure-data/td-skills
aps-doc-core
Core documentation generation patterns and framework for Treasure Data pipeline layers. Provides shared templates, quality validation, testing framework, and Confluence integration used by all layer-specific documentation skills.
treasure-data/td-skills
aps-doc-id-unification
Expert documentation generation for ID unification layers. Documents identity resolution algorithms, merge strategies, match rules, entity graphs, and multi-workflow orchestration. Use when documenting ID unification processes.
Didn't find tool you were looking for?