ReAct

What It Is

Reason + Act. The model interleaves thoughts, tool calls and observations. Introduced in Yao et al., 2022, "ReAct: Synergizing Reasoning and Acting in Language Models" (arXiv:2210.03629). The pattern behind most tool-using agents.

When to Use It

Use when the task needs tools, search or code execution. In modern APIs, prefer native tool use over hand-parsed Thought/Action text.

Blank Template

You can use these tools: {{tool list with descriptions}}
Loop: Thought -> Action -> Observation, until you can answer.
Stop and ask the user if {{escalation condition}}.
Final: {{answer format}}

Five Sample Prompts for Gen AI Professionals

REACT-01 - Research agent

Tools: web_search(query), fetch(url), notes_write(text).
Goal: Produce a sourced comparison of {{topic_options}} as of today.
Loop Thought -> Action -> Observation. Verify every claim against a fetched page.
Stop after {{max_steps}} steps or when every table cell has a source.
Final: Markdown table plus source list.

REACT-02 - Repo bug hunt

Tools: read_file(path), grep(pattern), run_tests(path).
Goal: Find why test {{test_name}} fails.
Loop Thought -> Action -> Observation. Never edit files.
Stop and report if the cause is outside the repo.
Final: root cause, file and line, proposed patch as a diff.

REACT-03 - RAG pipeline debugger

Tools: retrieve(query, k), show_chunk(id), embed(text).
Goal: Explain why {{query}} returns irrelevant chunks.
Loop Thought -> Action -> Observation. Test at least 3 hypotheses: chunking, embedding mismatch, query phrasing.
Final: evidence per hypothesis and the fix ranked by expected impact.

REACT-04 - Data analysis agent

Tools: python(code), read_csv(path).
Goal: Answer {{question}} from {{csv_path}}.
Loop Thought -> Action -> Observation. Every number in the answer must come from executed code.
Final: answer, the code used, and one chart description.

REACT-05 - MCP server smoke test

Tools: list_tools(), call_tool(name, args).
Goal: Verify every tool on {{server_name}} works.
Loop Thought -> Action -> Observation. Call each tool with a minimal valid input and one invalid input.
Final: table of tool, valid result, invalid result, pass/fail.

Links

• Up: Frameworks
• Curated by: curation prompts