6 techniques drawn from the 447 in our corpus tagged to this work,
each with the paper it came from. Across the techniques below the recurring targets are answers that are close but wrong, the same prompt giving different answers, output that ignores the requested format.
The techniques
01
Self-Refine / Self-Fix
Exploring Iterative Controllable Summarization with Large Language Models, arXiv:2411.12460
A two-stage iterative method for correcting LLM answers. First, the model analyzes and explains its error in not following numerical constraints, then generates a corrected version taking into account the analysis conducted. Increases accuracy of following formal requirements in 1-3 attempts.
Use it whenUse this technique when LLM ignores precise numerical constraints when generating text — for example, does not follow a specified resume length, word count, or number of sentences. Especially effective for controlled summarization tasks and managed content generation where accuracy in meeting formal requirements is important. The method is suitable for situations where the model well understands the task meaning but systematically violates quantitative parameters. Apply when you are ready to spend 1-3 iterations for d
Prompt
You are an expert summarizer tasked with condensing a 40-page report into a concise summary. Your primary challenge is to retain all critical numerical data and specific figures accurately while ensuring the summary is significantly shorter than the original report.
You will employ the Guide-to-Explain (GTE) iterative refinement technique to achieve this.
**Phase 1: Initial Summary Generation**
**Task:**
Summarize the provided 40-page report.
**Key Constraint:** The summary MUST include all specific numbers, statistics, percentages, dates, and figures present in the original report. Do not omit any quantitative data.
**(Imagine the LLM generates an initial summary here, likely failing to meet the strict numerical constraint.)**
---
**Phase 2: Iterative Refinement (GTE Method)**
**Instruction:**
Your initial summary did not fully adhere to the constraint of retaining *all* numerical data.
**Step A: Identify and Explain the Error**
1. **Compare:** Carefully compare your generated summary against the original 40-page report.
2. **Identify Missing Data:** Pinpoint at least two specific instances where numerical data (numbers, statistics, percentages, dates, figures) from the original report was omitted or misrepresented in your summary.
3. **Explain the Error:** For each identified instance, explain *why* you failed to include or accurately represent this specific piece of numerical data. Analyze the nature of the oversight (e.g., did you prioritize brevity over accuracy, did you misinterpret the data's importance, etc.?).
**Step B: Re-generate with Correction**
1. **Incorporate Analysis:** Based on your explanation of the errors, generate a *new*, revised summary.
2. **Strict Adherence:** This revised summary MUST now strictly include *every single piece of numerical data* found in the original 40-page report. Prioritize the accurate inclusion of all quantitative information.
3. **Conciseness:** While accuracy is paramount, aim for the summary to be as concise as possible *without sacrificing any numerical data*.
**Output Format:**
Present your response clearly, first showing the explanation of errors (Step A), followed by the revised, numerically accurate summary (Step B). Use bullet points for clarity in explaining errors.
Order Matters: Investigate the Position Bias in Multi-constraint Instruction Following, arXiv:2502.17204
A technique for ordering instructions in a prompt following the "hard-to-easy" principle. Complex constraints (word prohibitions, structure) are placed at the beginning of the instruction list, simple ones (capitalization, final phrase)—at the end. This forces the model to immediately account for difficult requirements during generation, improving accuracy of meeting all conditions.
Use it whenUse this technique when a prompt contains multiple instructions and constraints, especially if the model fails to execute them simultaneously. Ideal for tasks with complex requirements for text format, style and content. Apply when creating content with strict constraints (word prohibitions, character limits, specific structure). Especially effective for generating text where compliance with all conditions simultaneously is important. The method is universal and works with any LLM without
Prompt
**Role:** You are an AI assistant tasked with summarizing a 40-page report for a manager.
**Context:** The report contains detailed numerical data and findings. The manager needs a concise summary that retains all critical numbers and their significance.
**<INSTRUCTIONS>**
Follow these instructions, ordered from most complex to least complex, to ensure accuracy and adherence to constraints:
1. **[Very Complex] Numerical Data Extraction and Synthesis:** Identify and extract all specific numerical data points, statistics, percentages, and figures from the 40-page report. For each number, provide a brief explanation of its context and significance within the report's findings. Present this as a structured table with columns: "Metric", "Value", "Context/Significance".
2. **[Complex] Key Findings Condensation:** Distill the main conclusions and arguments of the report into a maximum of three bullet points. These points must directly relate to or be supported by the numerical data extracted in the previous step.
3. **[Medium] Overall Narrative Flow:** Ensure the summary flows logically, connecting the numerical data and key findings into a coherent narrative that a manager can easily understand.
4. **[Simple] Language and Tone:** Use clear, professional, and objective language suitable for a business report summary. Avoid jargon where possible, or explain it briefly if essential.
5. **[Very Simple] Output Format:** Begin the entire response with the title: "Executive Summary: [Report Title - infer from context if possible, otherwise use '40-Page Report']".
****
Begin your summary generation now.
Explain-then-Process: Using Grammar Prompting to Enhance Grammatical Acceptability Judgments, arXiv:2506.02302
A two-step method: first, the model explains the necessary rule or standard, then uses its own explanation as an explicit instruction for performing the main task. Converts implicit model knowledge into active directives, shifting focus from semantics to structure and rules.
Use it whenUse this technique when the model must strictly follow specific rules, standards, or style but ignores structural details in favor of general meaning. Particularly effective for tasks of assessing grammatical correctness, format compliance, applying stylistic norms, and verifying regulatory adherence. Suitable when the model knows the rule but doesn't apply it consistently. The method converts implicit knowledge into explicit instructions, forcing the model to first formulate the rule.
What Makes a Good Natural Language Prompt?, arXiv:2506.06950
A structured framework for creating and improving prompts based on 21 properties in 6 categories: communication, cognitive load, instructions, logic, factuality, and responsibility. Key principle: focusing on improving one correctly chosen property is more effective than simultaneously improving multiple aspects.
Use it whenUse this approach when you want to systematically improve prompt quality instead of intuitive formulation selection. Particularly useful when working on complex tasks where results are unstable or unsatisfactory. Apply when you need to diagnose weak points in a prompt and purposefully strengthen them. Ideal for creating production prompts where predictability and quality matter. Helps avoid overloading prompts with excessive techniques and focus on key improvements.
Self-Critique-Guided Curiosity Refinement: Enhancing Honesty and Helpfulness in Large Language Models via In-Context Learning, arXiv:2506.16064
Three-step prompting method: the model first generates an initial answer, then acts as a critic of its own answer by honesty and usefulness criteria, and finally creates an improved version based on self-criticism. Does not require model retraining, works through standard prompt chains.
Use it whenUse this technique when you need to get maximally honest, accurate, and helpful answers from a language model, especially in situations where information reliability matters. The method is particularly effective for complex questions requiring acknowledgment of model limitations, for tasks needing specific actionable recommendations, and when initial answers may contain inaccuracies or hallucinations. Suitable for research tasks, consulting, content creation, and any scenarios where answer quality is critical
ExpertPrompting: Instructing Large Language Models to be Distinguished Experts, arXiv:2305.14688
A method of improving LLM response quality by providing a detailed description of a specialized expert's role in the prompt. Before the main question, a description of the 'personality' of a specialist with a specific specialization, skills, and experience is given. This prompts the model to generate more structured, in-depth, and professional responses.
Use it whenUse Expert Prompting when you need more in-depth, detailed, and professional responses from the language model. Especially effective for specialized tasks requiring expert knowledge: writing technical documentation, creating marketing strategies, medical consultations, legal issues. Apply when standard queries provide superficial answers, and you need expert depth. Suitable for one-time queries and does not require complex setup — just add an expert role description before the main question.
A missing brief. No structure around a request invents the context the model
does not have: your constraints, your audience, your prior decisions. Technique work pays off
after the brief is right, not instead of it.
Stale tricks. Persona lines, politeness and offers of a reward have been
measured repeatedly through 2025 and 2026 and come out close to noise. Anything selling you
those is selling 2023.
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 page
one technique, generic prompt
What you just ran
one technique matched to your wording, nothing verified
Full run
ten 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.