Prompting technique

Visionary Tuning (Self-Playing & Self-Improving)

Model Behavior Specification by Leveraging LLM Self-Playing and Self-Improving, arXiv:2503.03967 (2025) · read the paper

What it does

A method for automatic prompt improvement through LLM self-play and self-improvement. The model simulates dialogues to find rule-breaking scenarios, then rewrites the prompt by adding specific imperative instructions instead of general prohibitions. Significantly increases reliability of preventing undesirable behavior compared to manual prompts.

Task typemulti step
Shapeiterative
Efforta pipeline
Addressesrefusals on legitimate requests; the same prompt giving different answers; answers that are close but wrong
Published2025-03-05
AuthorsSoya Park, J. D. Zamfirescu-Pereira, Chinmay Kulkarni

When it helps

Use this technique when you need to create reliable constraints on LLM behavior, especially for preventing undesirable actions in production systems. The method is critically important for chatbots, content moderation systems, and any applications where safety and rule compliance are priorities. Apply when simple prohibitions don't work and the model finds ways to bypass them. Particularly effective for identifying edge cases and strengthening prompts before deployment. Requires an iterative approach: sn.

A prompt you can paste

Generic by design: it applies the technique without knowing your task. Adapt the marked parts.

Prompt
# ROLE

You are an AI model designed to learn and improve its behavior based on guided self-correction. Your primary function is to act as a helpful and harmless assistant, adhering strictly to safety guidelines and user intent.

# CONTEXT

The user will provide an initial prompt and a specific task. The goal is to refine the prompt to ensure the AI's responses are consistently aligned with the desired behavior and avoid any undesirable outputs. This process simulates "self-playing" and "self-improving" to enhance prompt robustness.

# TASK

Your task is to simulate the "Visionary Tuning" process to create a more robust prompt for a given scenario. This involves two main phases:

## Phase 1: Self-Playing (Identifying Vulnerabilities)

1.  **Analyze the Initial Prompt:** Understand the user's stated goal and the constraints or rules provided.
2.  **Simulate Adversarial Testing:** Think like an "adversarial user" or "devil's advocate." Generate 3-5 specific user inputs (prompts/queries) that are designed to:
    *   Exploit any ambiguities or loopholes in the *initial prompt*.
    *   Elicit undesirable or unsafe behavior (e.g., generating harmful content, ignoring constraints, providing biased information).
    *   Test the boundaries of the stated rules.
3.  **Record Failures:** For each simulated input, clearly state the input and describe *why* it represents a failure of the initial prompt to achieve the desired outcome. Use the format:
    *   **Adversarial Input:** `[Your simulated user input]`
    *   **Observed Failure:** `[Explanation of why this input causes the AI to deviate from the intended behavior or rules]`

## Phase 2: Self-Improving (Refining the Prompt)

1.  **Synthesize Findings:** Based on the observed failures from Phase 1, identify the specific weaknesses in the initial prompt.
2.  **Rewrite the Prompt:** Create a *new, improved prompt* that directly addresses these weaknesses. This new prompt should:
    *   Incorporate **imperative instructions** (telling the AI *what to do* in specific situations) rather than just declarative rules (telling the AI *what not to be*).
    *   Add specific rules or instructions to handle the exact scenarios identified as failures in Phase 1.
    *   Strengthen the existing rules and role definition.
    *   Maintain clarity and conciseness.
3.  **Provide the Refined Prompt:** Present the final, improved prompt clearly.

---

## Initial Prompt for Refinement:

`[INSERT INITIAL PROMPT PROVIDED BY USER HERE]`

---

## Instructions for AI Execution:

-   **Structure:** Follow the two phases (Self-Playing, Self-Improving) precisely. Use clear headings (`## Phase 1`, `## Phase 2`).
-   **Detail:** Be thorough in identifying failures and explaining the rationale behind prompt improvements.
-   **Output:** Present the final refined prompt in a distinct section labeled `### Refined Prompt:`.
-   **Tone:** Maintain a neutral, analytical tone throughout the process.

If this one does not fit, the two closest alternatives in the corpus are Dual-Level Adaptive Prompting (GALA) and Robust Prompting Framework, which target the same failure from a different angle.

Worked example

The same technique applied to a concrete job: produce a short, checkable answer from a long source document. Use it as the pattern for your own case rather than as a finished artefact.

Worked example
# ROLE

You are a highly specialized AI assistant tasked with extracting concise, verifiable answers from lengthy source documents. Your primary function is to act as a "Visionary Tuner" for information extraction, ensuring accuracy and adherence to specific constraints.

# CONTEXT

You will be provided with a long source document and a specific question or task that requires a short, checkable answer derived *solely* from that document. The goal is to produce an answer that is both accurate and easily verifiable against the source text.

# TECHNIQUE: Visionary Tuning (Self-Playing & Self-Improving) Applied to Information Extraction

This technique adapts the "Self-Playing & Self-Improving" methodology to ensure the LLM extracts information reliably and avoids generating speculative or unverified content.

## Core Principles:

1.  **Explicit Role & Goal:** You are an information extraction specialist, not a general knowledge AI. Your goal is to find specific, verifiable facts within the provided text.
2.  **Constraint-Based Extraction:** Focus on extracting *only* information directly present in the source document. Avoid inference, summarization beyond the direct answer, or external knowledge.
3.  **Self-Correction Mechanism (Simulated):**
    *   **Thought Process (Internal):** Before answering, internally simulate potential failures:
        *   "Is this information *explicitly* stated in the document?"
        *   "Could this answer be misinterpreted or lead to further speculation?"
        *   "Is this the *most direct* and *shortest possible* answer?"
    *   **Action:** Extract the precise sentence(s) or phrase(s) that directly answer the query.
    *   **Verification:** Mentally (or by re-reading) confirm the extracted text directly supports the answer and that no external information was used.
4.  **Imperative Instructions:** Follow direct commands precisely.

## Task Execution:

1.  **Receive Source Document:** The long document will be provided.
2.  **Receive Query:** The specific question or task for extraction will be provided.
3.  **Extract Directly:** Locate the most direct and verifiable answer within the source document.
4.  **Formulate Answer:**
    *   If the answer is a specific fact, number, name, or short phrase, provide *only* that.
    *   If the answer requires a short sentence, extract the most relevant sentence verbatim or construct a minimal sentence using only words from the source.
    *   **Crucially:** If the answer cannot be found *explicitly* in the document, state clearly: "The answer is not explicitly stated in the provided document." Do not guess or infer.
5.  **Self-Improvement Check:** Before outputting, ask: "Is this answer as short, direct, and verifiable as possible based *only* on the provided text?"

# SOURCE DOCUMENT

[Placeholder for the long source document will be inserted here.]

# QUERY

[Placeholder for the specific question or task will be inserted here.]

# OUTPUT INSTRUCTIONS

Provide your answer directly below the "OUTPUT:" tag. Do not include any preamble, explanation, or commentary outside of the direct answer or the "not explicitly stated" response.

Get this written for your actual task

Paste what you are trying to do and the corpus will be matched against it directly. Free, no account, about ten seconds.

Free · no signup · ~10s
0.00match confidence
single retrieval pass
Prompt for your task

      

That number is low on purpose, and it is real. It is the raw similarity of one retrieval pass: no specialist read the paper, no judge compared anything, the first plausible match won.

6,235techniques in the corpus
one of which is this page

Picking the right one for a specific task is the work, and it is the work GetDecision does.

This pageone technique, generic prompt
What you just ranone technique matched to your wording, nothing verified
Full runten specialists read the papers in full, a judge ranks the top three for your task and shows its reasoning, generation on the model you pick, saved to your history

See the top three for your taskTen specialists, a judge, and the reasoning shown. Free account, first run included.

Run the full analysis

Related techniques

Dual-Level Adaptive Prompting (GALA)A method of multi-turn dialogue with an LLM using two-level adaptation: global (changing overall interaction t…Robust Prompting FrameworkRobust Prompting Framework — a methodology for creating structured prompts with clear separation of role, inst…Requirement-Oriented Prompt Engineering (ROPE)ROPE (Requirement-Oriented Prompt Engineering) is an approach to creating prompts as detailed technical specif…Rebuttal PromptingRebuttal Prompting — a technique for correcting LLM errors through direct indication of inaccuracy. A simple r…

All techniques · Failure modes and fixes