6 techniques drawn from the 369 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 model losing what was agreed earlier, the same prompt giving different answers.
The techniques
01
Language of Thoughts (LoT)
On the Thinking-Language Modeling Gap in Large Language Models, arXiv:2505.12896
A technique that forces the model to first explicitly extract and rephrase key information from context (expand and echo commands), organizing everything systematically, and only then give the final answer. Solves the problem of the model getting lost in complex context through two-stage processing: information structuring and then reasoning.
Use it whenApply LoT when the model has access to necessary knowledge but gets lost in complex or voluminous context. Especially effective when dealing with confusing formulations, when key information is hidden among irrelevant details, or when you need to extract and structure facts from large texts. Ideal for document analysis, legal texts, research materials where interpretation accuracy is critical. Use when reasoning quality matters more than response speed.
Prompt
You are an expert legal analyst and strategist. Your task is to analyze a complex legal scenario and propose a multi-step strategic plan, applying the Language of Thoughts (LoT) methodology.
**Scenario:** A client is involved in a contract dispute. The opposing party has breached the contract by failing to deliver goods on time and providing substandard quality. The contract includes a dispute resolution clause that mandates arbitration before litigation. The client wants to recover damages for the breach, including lost profits due to the delay and the cost of rectifying the substandard goods.
**Task:** Develop a strategic plan for the client.
**Instructions:**
Before providing the final strategic plan, you must perform the following preparatory steps based on the LoT technique:
1. **Expand (L-implicitness):** For each of the following points, expand and clarify the implications:
* "Contract breach by failing to deliver goods on time"
* "Substandard quality of delivered goods"
* "Mandatory arbitration clause before litigation"
* "Client seeks recovery of lost profits and cost of rectification"
2. **Echo (Q-implicitness):** Identify and list all relevant facts from the scenario that are crucial for developing the strategy. Exclude any extraneous details.
* List all parties involved.
* List all contractual obligations and breaches.
* List all remedies sought by the client.
* List all procedural constraints (e.g., arbitration).
After completing the "Expand" and "Echo" steps, and based *only* on the clarified information and extracted facts, provide a comprehensive, multi-step strategic plan. This plan should detail the recommended course of action, including:
* Initial steps to take regarding the arbitration process.
* Evidence to gather to support claims of delay and substandard quality.
* Methods for quantifying lost profits and rectification costs.
* Potential negotiation strategies.
* Contingency plans if arbitration does not yield a satisfactory outcome.
**Output Format:**
Structure your response with clear headings for each section:
## Preparatory Steps
### 1. Expand (L-implicitness)
### 2. Echo (Q-implicitness)
## Strategic Plan
[Your detailed multi-step plan here]
**Ensure your plan is actionable and directly addresses the client's objectives within the contractual framework.**
Inference Scaling for Long-Context Retrieval Augmented…
Inference Scaling for Long-Context Retrieval Augmented Generation, arXiv:2410.04343
Method for improving RAG systems through model training on examples (DRAG) and iterative decomposition of complex questions (IterDRAG). Demonstrates to the model how to work with documents by showing question-context-answer patterns, then breaks complex queries into simple subtasks with separate searches for each. Quality increases linearly with effort.
Use it whenApply this technique when standard RAG shows low accuracy or the model ignores provided context. Particularly effective for complex questions requiring information synthesis from multiple sources, comparative analysis, or multi-step reasoning. Ideal for working with large knowledge bases, legal documents, research tasks. DRAG is suitable when you need to teach the model to use context correctly through examples. IterDRAG is necessary for questions that can be broken down into subtasks, where each step requires separate information search. Requires computational resources proportional to the number of iterations.
Prompt
You are an expert legal analyst specializing in multi-step planning and information retrieval. Your task is to simulate the application of the DRAG (Demonstration-based RAG) and IterDRAG methodology to a complex legal planning problem.
**Scenario:** A client needs to plan a strategy for a new business venture that involves navigating complex regulatory requirements and potential market challenges.
**Objective:** Develop a comprehensive, multi-stage strategic plan by leveraging provided information and demonstrating a structured approach to information retrieval and analysis.
**Methodology:** Apply the principles of DRAG and IterDRAG. This involves:
1. **Demonstration (DRAG):** Providing clear examples of how to break down a complex question, retrieve relevant context, and synthesize an answer.
2. **Iterative Decomposition (IterDRAG):** Structuring the problem into sequential, manageable steps, where the output of one step informs the next.
---
**[DRAG EXAMPLE 1: Regulatory Analysis]**
**Question:** What are the primary regulatory hurdles for a tech startup offering AI-driven personalized financial advice in California?
**Context:**
* California Consumer Privacy Act (CCPA) / California Privacy Rights Act (CPRA)
* Securities and Exchange Commission (SEC) regulations on investment advice
* Financial Industry Regulatory Authority (FINRA) rules
* State-specific financial licensing requirements
* Data security and breach notification laws
**Answer:**
The primary regulatory hurdles include:
1. **Data Privacy:** Compliance with CCPA/CPRA regarding the collection, use, and protection of personal financial data.
2. **Investment Advisor Registration:** Potential requirement to register as an investment advisor with the SEC and/or state securities regulators, depending on the specific services offered.
3. **FINRA Compliance:** Adherence to FINRA rules if the service is deemed to involve securities transactions or broker-dealer activities.
4. **State Licensing:** Obtaining necessary financial services licenses in California and potentially other states where users are located.
5. **Data Security:** Implementing robust security measures to prevent breaches and complying with notification requirements.
---
**[DRAG EXAMPLE 2: Market Entry Strategy]**
**Question:** Outline a phased market entry strategy for a new AI financial advisor targeting young professionals (25-35) in major US tech hubs.
**Context:**
* Target demographic: Young professionals, tech-savvy, value convenience and personalization.
* Competitive landscape: Established robo-advisors, traditional wealth managers, emerging fintech apps.
* Key differentiators: Advanced AI personalization, focus on long-term financial planning for career growth.
* Phased approach: Beta launch, regional rollout, national expansion.
**Answer:**
A phased market entry strategy:
1. **Phase 1: Beta Launch (3-6 months):**
* **Target:** Limited group of early adopters in 1-2 key tech hubs (e.g., San Francisco, Austin).
* **Objective:** Gather user feedback, refine AI algorithms, test core features, build initial testimonials.
* **Activities:** Private beta access, intensive user support, focus on product-market fit.
2. **Phase 2: Regional Rollout (6-12 months):**
* **Target:** Expand to 3-5 additional major tech hubs.
* **Objective:** Increase user acquisition, optimize marketing channels, establish brand presence.
* **Activities:** Targeted digital marketing, strategic partnerships with employers, PR.
3. **Phase 3: National Expansion (12+ months):**
* **Target:** All major US markets.
* **Objective:** Achieve significant market share, scale operations, introduce advanced features.
* **Activities:** Broad marketing campaigns, potential M&A, continuous product innovation.
---
**[YOUR TASK: COMPREHENSIVE LEGAL STRATEGY PLAN]**
**Overall Objective:** Develop a strategic plan for a new AI-driven fintech company providing personalized financial advice. The company aims to serve young professionals (25-35) in California initially, with plans for national expansion. Key considerations include regulatory compliance, market positioning, and operational scaling.
**Provided Context:**
* **Company Offering:** AI platform for personalized financial planning, investment recommendations, and budgeting tools.
* **Target Audience:** Young professionals (25-35) in California, tech-savvy, value convenience, seeking long-term financial growth.
* **Initial Market:** California.
* **Expansion Goal:** National US market within 3 years.
* **Key Differentiators:** Advanced AI personalization, user-friendly interface, focus on career-related financial goals.
* **Regulatory Landscape (General):** Requires adherence to financial advisory laws, data privacy regulations (CCPA/CPRA), and consumer protection laws. Specific state and federal registrations may be necessary.
* **Competitive Landscape:** Mix of established robo-advisors, traditional banks, and newer fintech apps.
**Instructions:**
Structure your response as a multi-stage strategic plan, applying the principles of DRAG and IterDRAG.
---
**Phase 1: Foundational Regulatory & Compliance Strategy (Apply DRAG principles)**
* **Step 1.1: Identify Key Regulatory Areas.** Based on the provided context and general knowledge of fintech, list the *critical* regulatory domains the company *must* address for its California launch. Use the DRAG example format (Question, Context, Answer) to structure this.
* **Question:** What are the essential regulatory domains for an AI financial advisor targeting young professionals in California?
* **Context:** [Use the "Provided Context" above, focusing on company offering, target audience, initial market, and general regulatory landscape.]
* **Answer:** [Provide a structured answer, similar to Example 1, listing 4-6 key regulatory areas.]
* **Step 1.2: Define Initial Compliance Actions.** For each identified regulatory area, list 1-2 concrete, actionable steps the company must take *before* launching in California.
* **Question:** What are the immediate compliance actions required for launching an AI financial advisor in California?
* **Context:** [Refer to the identified regulatory areas from Step 1.1 and the provided context.]
* **Answer:** [List specific actions, e.g., "Consult with legal counsel specializing in securities law," "Develop a comprehensive CCPA/CPRA compliance policy."]
---
**Phase 2: Market Entry & Product Development Strategy (Apply IterDRAG principles)**
* **Step 2.1: Phased Market Entry Plan.** Decompose the market entry into at least three distinct phases (e.g., Beta, Regional, National). For each phase, define:
* **Objective:** What is the primary goal of this phase?
* **Target Audience/Market:** Who are you reaching?
* **Key Activities:** What actions will be taken?
* **Success Metrics:** How will success be measured?
* *Use a format similar to DRAG Example 2.*
* **Step 2.2: Feature Prioritization for Launch.** Based on the target audience (young professionals) and differentiators (AI personalization, user-friendliness, career focus), identify and prioritize the top 5 core features for the initial launch (Phase 1). Justify each prioritization.
* **Question:** What are the top 5 essential features for the initial launch of an AI financial advisor targeting young professionals?
* **Context:** [Company Offering, Target Audience, Key Differentiators, Phase 1 Objectives.]
* **Answer:** [List 5 features with brief justifications linking them to audience needs and company goals.]
---
**Phase 3: Risk Assessment & Mitigation**
* **Step 3.1: Identify Key Risks.** Identify 3-4 significant risks associated with this venture, considering both regulatory and market aspects.
* **Question:** What are the most critical risks for a new AI financial advisor startup in California?
* **Context:** [All previously provided context, including regulatory landscape, competitive landscape, and proposed strategy.]
* **Answer:** [List 3-4 risks, e.g., "Failure to obtain necessary licenses," "Intense competition leading to slow user acquisition," "Data breach impacting user trust."]
* **Step 3.2: Propose Mitigation Strategies.** For each identified risk, suggest at least one concrete mitigation strategy.
* **Question:** How can the identified risks for the AI financial advisor startup be mitigated?
* **Context:** [Refer to the identified risks from Step 3.1.]
* **Answer:** [Provide specific mitigation strategies for each risk.]
---
**Final Synthesis:**
Provide a brief concluding statement summarizing the overall strategic direction, emphasizing the importance of a phased, compliance-first approach.
Cognitive DebiasingLargeLanguageModelsfor Decision Making
Self-Adaptive Cognitive Debiasing for Large Language Models in Decision-Making, arXiv:2504.04141
A three-step self-checking method for LLMs before generating answers. The model first analyzes the request for cognitive biases, then rewrites it in neutral form, and only then provides the final answer. Allows obtaining more objective results in decision-making tasks.
Use it whenApply this technique when working on decision-making tasks in critical areas (finance, medicine, law) where objectivity and absence of cognitive biases are important. Especially effective when the original request may contain emotional coloring, stereotypes or bias (herd behavior effect, loss aversion, information availability). Use to obtain more balanced recommendations in business analytics, strategic planning, risk assessment and any situations where the cost of error is high
Read Before You Think: Mitigating LLM Comprehension Failures with Step-by-Step Reading, arXiv:2504.09402
Technique of two-pass prompt processing: the model first breaks down the request into parts and identifies all conditions, then reconsiders each part taking into account subsequent requirements. Solves the problem of LLM ignoring conditions from the end of the prompt that change the meaning of the beginning. Increases accuracy from 67% to 89% on tasks with contradictory conditions.
Use it whenApply SSR++ when a prompt contains multiple conditions, especially if later requirements contradict or clarify earlier parts of the instruction. Ideal for complex business tasks with multi-level constraints: drafting documents with special requirements, planning with exceptions, content generation where final conditions are critical. Use when you notice the model ignoring important conditions from the end of the prompt. Particularly effective for tasks where a person would reread the instruction twice themselves
"Lost-in-the-Later": Framework for Quantifying Contextual Grounding in Large Language Models, arXiv:2507.05424
A technique to combat the 'lost in the end' effect, when LLM ignores information at the end of long prompts. Two instructions in the prompt force the model to process all context uniformly: 'use only this text' and 'read all text uniformly'. Reduces hallucinations and improves answer accuracy.
Use it whenApply this technique when working with long documents, reports, or contexts where information from all parts of the text is critically important, not just the beginning. Especially effective for analyzing legal documents, scientific articles, technical documentation where key details can be anywhere. Use when creating RAG systems for uniform information extraction from large fragments. Necessary when you notice the model ignoring data from the end of the prompt or making up
CLAP: Coreference-Linked Augmentation for Passage Retrieval, arXiv:2508.06941
Two-stage method for preprocessing long texts to improve information retrieval. LLM first splits the document into semantic blocks, then replaces all pronouns with specific nouns, making each fragment self-contained. Increases information extraction accuracy by 15-25% in question-answering tasks.
Increases information extraction accuracy by 15-25% in question-answering tasks.checked against the paper
Use it whenUse CLAP when working with long documents (articles, reports, legal documents, correspondence logs) where information extraction accuracy and context preservation are important. Especially effective for RAG systems and question-answering applications where text contains many pronouns and coreferences. Apply when standard text segmentation into fragments leads to loss of semantic connections and ambiguity. Ideal for building knowledge bases, analytical systems, legal document analysis...
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.