ReAct Pattern Skill

Bonded to: react-pattern-agent

---

Quick Start

Skill("custom-plugin-prompt-engineering:react-pattern")

---

Parameter Schema

parameters:
  max_iterations:
    type: integer
    default: 10
    description: Maximum reasoning-action cycles

  tool_format:
    type: enum
    values: [openai_function, anthropic_tool, langchain, custom]
    default: openai_function

  termination:
    type: enum
    values: [final_answer, max_iterations, tool_signal]
    default: final_answer

---

Core Pattern

You are an AI assistant that uses tools to accomplish tasks.

## Available Tools
[Tool definitions with descriptions and parameters]

## Response Format
For each step, use exactly this format:

Thought: [Your reasoning about what to do next]
Action: [tool_name({"param": "value"})]
Observation: [Result from the tool - provided by system]

Continue until you can provide:
Thought: I now have enough information to answer.
Final Answer: [Your complete response]

## Rules
1. Always think before acting
2. Use exactly one tool per action
3. Base decisions on observations
4. Never fabricate tool results
5. Stop when you have sufficient information

---

Component Definitions

Component	Purpose	Format
Thought	Reasoning about current state	Thought: I need to...
Action	Tool invocation	Action: tool({"param": value})
Observation	Tool result (system-provided)	Observation: {result}
Final Answer	Task completion	Final Answer: [response]

---

Tool Definition Schema

tool_schema:
  name:
    type: string
    description: "Unique identifier for the tool"

  description:
    type: string
    description: "When and why to use this tool"

  parameters:
    type: object
    properties:
      param_name:
        type: string|number|boolean|array|object
        description: "What this parameter does"
        required: true|false

  returns:
    type: string
    description: "What the tool returns"

example:
  name: search
  description: "Search the web for current information"
  parameters:
    query:
      type: string
      description: "Search query"
      required: true
    limit:
      type: integer
      description: "Maximum results"
      required: false
      default: 5
  returns: "List of search results with titles and snippets"

---

Workflow Patterns

Linear (Sequential)

Thought → Action → Observation → Thought → Action → Observation → Final Answer

Branching (Conditional)

Thought → Action → Observation
                       ↓
            ┌─ if condition A → Path 1
            └─ if condition B → Path 2

Iterative (Refinement)

Thought → Action → Observation
    ↑                    ↓
    └──── not satisfied ─┘
              ↓
         satisfied
              ↓
        Final Answer

---

Error Handling

error_strategies:
  tool_not_found:
    detection: "Invalid tool name in action"
    response: "List available tools and retry"

  invalid_parameters:
    detection: "Parameter validation fails"
    response: "Show correct schema, ask to retry"

  tool_timeout:
    detection: "No response within timeout"
    response: "Retry once, then proceed without"

  infinite_loop:
    detection: "Exceeds max_iterations"
    response: "Force Final Answer with partial results"

  all_tools_failed:
    detection: "All attempted tools return errors"
    response: "Respond without tools, acknowledge limitation"

---

Example Implementation

User: What's the weather in Tokyo and should I bring an umbrella?

Thought: I need to find the current weather in Tokyo. I'll use the weather tool.
Action: get_weather({"location": "Tokyo, Japan"})
Observation: {"temperature": 18, "condition": "cloudy", "precipitation_chance": 75%, "forecast": "Rain expected this afternoon"}

Thought: The precipitation chance is 75% with rain expected. I should recommend bringing an umbrella.
Final Answer: The weather in Tokyo is currently 18°C and cloudy. There's a 75% chance of precipitation with rain expected this afternoon. Yes, I recommend bringing an umbrella!

---

Validation

validation_checklist:
  tool_definitions:
    - [ ] All tools have clear descriptions
    - [ ] Parameters are well-documented
    - [ ] Return values are specified

  format:
    - [ ] Thought/Action/Observation format enforced
    - [ ] Termination conditions clear
    - [ ] Max iterations set

  behavior:
    - [ ] Always thinks before acting
    - [ ] Never fabricates observations
    - [ ] Handles errors gracefully

---

Troubleshooting

Issue	Cause	Solution
Never stops	No termination condition	Add max iterations
Wrong tool used	Vague tool descriptions	Make descriptions specific
Skips reasoning	Format not enforced	Add strict format check
Hallucinates results	No grounding	Require actual tool calls
Inefficient paths	No planning	Add planning step

---

Integration

integrates_with:
  - agent-design: Agent architecture
  - prompt-design: Base prompt structure
  - chain-of-thought: Reasoning patterns

tool_frameworks:
  - LangChain agents
  - OpenAI function calling
  - Anthropic tool use
  - Custom implementations

---

References

See references/GUIDE.md for advanced agent patterns. See assets/config.yaml for configuration options.