Prompt Engineering for Agentic Systems

Overview

Generate optimized prompts for agentic systems with clear rationale for technique selection.

Core Principles:

• Match technique to task - Different agent types require different prompting approaches
• Trade-offs matter - Always consider cost, latency, and accuracy when selecting techniques
• Structure over verbosity - Well-organized prompts outperform long unstructured ones
• Test and iterate - Verify prompts work before deploying to production

For broader agentic system design (choosing workflows vs agents, ACI/tool specifications, guardrails, multi-agent patterns), see the ai-engineering skill.

When to Use

Invoke this skill when:

• Creating prompts for tool-using agents (ReAct pattern)
• Designing prompts for planning or strategy agents
• Building prompts for data processing or validation agents
• Reducing hallucinations in fact-based tasks
• Optimizing prompt performance or cost

Common Scenarios

Scenario 1: Tool-Using Agent (ReAct)

Use when: Agent needs to reason and use tools autonomously

## SYSTEM
You are a research agent. Your goal is to gather information and synthesize findings.

## INSTRUCTIONS
Follow this pattern for each action:
Thought: [what you want to do]
Action: [tool name and parameters]
Observation: [result from tool]

When you have enough information, provide a final summary.

## AVAILABLE TOOLS
- search(query): Search for information
- read(url): Read a webpage
- finish(summary): Complete the task

## STOP CONDITION
Stop when you have answered the user's question or gathered sufficient information.

---

Scenario 2: Planning Agent (Tree of Thoughts)

Use when: Agent needs to explore multiple approaches before committing

## TASK
Design a migration strategy from monolith to microservices.

## INSTRUCTIONS
Generate 3 different approaches:

Approach 1: [description]
- Pros: [list]
- Cons: [list]
- Effort: [estimate]

Approach 2: [description]
- Pros: [list]
- Cons: [list]
- Effort: [estimate]

Approach 3: [description]
- Pros: [list]
- Cons: [list]
- Effort: [estimate]

Then select the best approach and explain your reasoning.

---

Scenario 3: Data Validation (Few-Shot with Negative Examples)

Use when: Agent needs consistent output format and should avoid common errors

## TASK
Validate email addresses and return structured JSON.

## VALID EXAMPLES
Input: user@example.com
Output: {"valid": true, "reason": "proper email format"}

Input: user.name@company.co.uk
Output: {"valid": true, "reason": "proper email format"}

## INVALID EXAMPLES (what NOT to accept)
Input: user@.com
Output: {"valid": false, "reason": "invalid domain format"}

Input: @example.com
Output: {"valid": false, "reason": "missing local part"}

Input: user example.com
Output: {"valid": false, "reason": "missing @ symbol"}

## NOW VALIDATE
Input: {user_input}

---

Scenario 4: Factual Accuracy (Chain-of-Verification)

Use when: Reducing hallucinations is critical

## TASK
Explain how transformers handle long-context windows

## STEP 1: Initial Answer
Provide your explanation...

## STEP 2: Verification Questions
Generate 5 questions that would expose errors in your answer:
1. [question]
2. [question]
3. [question]
4. [question]
5. [question]

## STEP 3: Answer Verification
Answer each verification question factually...

## STEP 4: Final Answer
Refine your original answer based on verification results...

---

Scenario 5: Complex Decision (Structured Thinking)

Use when: Agent needs to analyze trade-offs before deciding

## TASK
Recommend: microservices or monolith for our startup?

## THINKING PROTOCOL

[UNDERSTAND]
- Restate the problem in your own words
- Identify what's actually being asked

[ANALYZE]
- Break down into sub-components
- Note assumptions and constraints

[STRATEGIZE]
- Outline 2-3 approaches
- Evaluate trade-offs

[EXECUTE]
- Provide final recommendation
- Explain reasoning

---

Scenario 6: Self-Improving Output (Self-Refine)

Use when: You want the agent to review and improve its own work

## TASK
Write a README for the checkout API

## STEP 1: Initial Draft
[Generate initial README]

## STEP 2: Critique
Identify 3-5 improvements needed:
- [weakness 1]
- [weakness 2]
- [weakness 3]

## STEP 3: Refinement
Rewrite addressing all identified improvements...

---

Quick Decision Tree

Use this table to select techniques quickly:

Agent Characteristic	Recommended Technique
Uses tools autonomously	ReAct
Planning/strategy with alternatives	Tree of Thoughts
High-stakes correctness	Self-Consistency
Factual accuracy, hallucination reduction	Chain-of-Verification (CoVe)
Single-path complex reasoning	Chain of Thought
Complex decisions with trade-offs	Structured Thinking Protocol
Reducing bias, multiple viewpoints	Multi-Perspective Prompting
Uncertainty quantification	Confidence-Weighted Prompting
Proprietary documentation, prevent hallucinations	Context Injection with Boundaries
Self-review and improvement	Self-Refine
Breaking complex problems into subproblems	Least-to-Most Prompting
High-quality content through multiple passes	Iterative Refinement Loop
Multi-stage workflows with specialized prompts	Prompt Chaining
Improving recall for factual questions	Generated Knowledge Prompting
Unclear how to structure the prompt	Meta-Prompting (nuclear option)
Strict technical requirements	Constraint-First Prompting
Requires consistent format/tone	Few-Shot (supports negative examples)
Simple, well-defined task	Zero-Shot
Domain-specific expertise	Role Prompting
Procedural workflow	Instruction Tuning

For detailed decision logic with branching, see decision-tree.md

Technique Reference

All available techniques with examples, use cases, and risks: techniques.md

Critical Anti-Patterns

Common mistakes to avoid: anti-patterns.md

Critical warnings:

• Do NOT use ReAct without tools - Adds unnecessary complexity
• Do NOT use Tree of Thoughts for deterministic problems - Single correct answer doesn't need alternatives
• Do NOT use vague roles - "Expert" without scope provides little benefit
• Do NOT omit stop conditions - Agents may continue indefinitely
• Do NOT use Self-Refine for objective tasks - Calculations don't need self-critique

Canonical Template

Use this template as the foundation for generated prompts: template.md

Basic structure:

## SYSTEM / ROLE
You are a [specific role] with authority over [explicit scope]
Boundaries: [what you must NOT do]

## TASK
[Single clear goal in one sentence]

## INSTRUCTIONS
Follow these steps:
1. [First step]
2. [Second step]

Constraints:
- [Specific limits]
- [Format requirements]

## STOP CONDITION
Stop when: [success criteria]

Output Rationale Template

When delivering a generated prompt, use this structure:

## Generated Prompt for [Agent Name/Type]

[prompt in code block]

## Rationale

**Agent Type**: [Tool-using / Planner / Conversational / Data-processor]

**Task Complexity**: [Simple / Multi-step / Planning-heavy]

**Techniques Used**:
- [Technique]: [Why it works for this use case]

**Expected Behavior**: [What the agent will do]

**Trade-offs**: [Cost, latency, flexibility - ALWAYS include if technique increases tokens or latency]

**Considerations**: [Edge cases, limitations, or risks]

Guardrail Rule

If a prompt increases latency, token usage, or operational cost, this MUST be stated explicitly in the rationale under "Trade-offs."

Techniques that increase cost/latency:

• Self-Consistency (multiple generations)
• Chain-of-Verification (multiple passes)
• Iterative Refinement Loop (multiple passes)
• Self-Refine (multiple passes)
• Tree of Thoughts (exploring alternatives)
• Least-to-Most Prompting (sequential subproblems)