PROMPTWIK
Mobile_Node_Online
← Back to Techniques
Level: Master // Technical Analysis

Reflexion

An iterative self-correction framework where the AI agent receives feedback on previous failures, 'reflects' on the root cause, and retries the task with improved logic.

Software Engineers QA Testers Backend Developers
Production Use Case

Task Context

A coding agent generates a Python script that fails during execution. The error log is passed back to the agent for autonomous correction.

Input (Prompt) 
The script you generated failed with the following error: 'IndexError: list index out of range'. Reflect on what caused this error within your 'for' loop and generate the corrected code.
Output (AI Result) 
Reflexion: The loop 'for i in range(len(mylist) + 1)' attempts to access a non-existent index because Python lists are zero-indexed. I must remove the '+ 1'.

Corrected Code:
for i in range(len(mylist)):
  print(mylist[i])

What is the Reflexion Technique?

The Reflexion technique empowers Large Language Models (LLMs) with the capability for iterative self-correction. It consists of designing a Prompt Engineering loop where the AI generates a response, receives feedback indicating a failure, and is then instructed to “reflect” verbally on why it erred before attempting a new solution.

This verbalization of the error forces the model’s neural network to recalibrate its context window. It is the “secret sauce” that allows modern AI coding assistants (such as Cursor or GitHub Copilot) to exhibit near-human intelligence when autonomously debugging complex software failures.

When to Use Reflexion?

Reflexion excels in tasks where success can be objectively measured by an external system, a compiler, or a strict set of logical constraints.

  • Autonomous Debugging: The AI writes code, the interpreter (e.g., Python) throws an exception, and the reflexion prompt passes the traceback back to the model for correction.
  • Strict Format Enforcement: If the AI is tasked with generating a JSON object but omits a comma, an automated validator detects the error and triggers a Reflexion prompt for self-correction without human intervention.
  • Educational & Adaptive Tutoring: Models designed to gauge a user’s knowledge level, reflecting on and adjusting their pedagogical strategy if a user indicates they do not understand a specific concept.
  • Autonomous Agents: Allowing an AI web-scraper to recognize if it has navigated to the incorrect URL or encountered a CAPTWEA, subsequently reflecting on its search strategy to find an alternative path.

Technical Limitations & Trade-offs

As with other iterative strategies, Reflexion does not guarantee an eventual solution. In some scenarios, the Artificial Intelligence may enter a “Blind Feedback Loop,” where it reflects on its error but repeatedly commits the same logical fallacy.

Implementing this technique requires setting strict Iteration Limits (e.g., a maximum of 3 reflexions) to prevent excessive OpenAI or Anthropic API costs. Furthermore, the model used must be advanced enough (Claude 3.5+, GPT-4o, or O1) to possess the cognitive depth required for meaningful self-critique.