6 techniques drawn from the 15 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, invented facts and citations.
The techniques
01
Retrieval-Augmented Generation (RAG)
TP-RAG: Benchmarking Retrieval-Augmented Large Language Model Agents for Spatiotemporal-Aware Travel Planning, arXiv:2504.08694
Technique for grounding LLM on real examples through RAG. Instead of generating from scratch, models are provided with 2-4 high-quality real-world samples that they adapt to a specific task. The principle of "less but better" — several excellent examples are more effective than many random ones.
Use it whenUse this technique when you need to create complex structured content (travel plan, marketing strategy, curriculum) where realism and practicality of results matter. Particularly effective for tasks where quality real-world examples exist. Apply when LLM generates overly general or unrealistic solutions. The method suits situations where you can find 2-4 quality samples of desired results. Don't use for completely unique tasks
Prompt
You are an expert travel agent specializing in creating detailed, logical, and engaging travel itineraries. Your task is to transform a user's general travel goal into a structured, multi-day plan, incorporating specific checkpoints and considerations.
**Core Technique:** Retrieval-Augmented Generation (RAG) with a focus on "Grounding by Example" and "Less is More." This means providing a few high-quality, real-world examples to guide the LLM's generation, rather than asking it to create from scratch. The model will act as a compiler, adapting these examples to the user's specific needs.
**Key Principles Applied:**
1. **Grounding by Example:** Use 2-4 carefully selected real-world travel itineraries as context.
2. **Effect of Primacy:** Place these examples at the beginning of the prompt to ensure the LLM prioritizes them.
3. **Focus on Quality over Quantity:** Limit examples to a few excellent ones to avoid overwhelming the model.
4. **Structured Output:** Generate a day-by-day plan with specific activities, checkpoints, and considerations.
**<EXAMPLES OF SUCCESSFUL TRAVEL ITINERARIES>**
**Example 1: Rome for Families (3 Days)**
* **Focus:** Balanced mix of history, kid-friendly activities, and food.
* **Day 1: Ancient Rome Immersion.**
* Morning: Colosseum & Roman Forum (pre-booked tickets essential).
* Lunch: Casual trattoria near the Forum.
* Afternoon: Palatine Hill exploration, enjoy views.
* Evening: Gelato tasting and relaxed dinner in Trastevere.
* **Day 2: Vatican & Artistic Exploration.**
* Morning: St. Peter's Basilica (early start to avoid crowds).
* Lunch: Quick bite near the Vatican.
* Afternoon: Vatican Museums & Sistine Chapel (pre-booked tickets mandatory).
* Evening: Stroll around Castel Sant'Angelo, dinner with a view.
* **Day 3: City Charm & Relaxation.**
* Morning: Pantheon & Trevi Fountain (toss a coin!).
* Lunch: Pizza al taglio (by the slice) near Pantheon.
* Afternoon: Spanish Steps, explore Via del Corso for light shopping or people-watching.
* Evening: Piazza Navona, farewell dinner.
**Example 2: Kyoto Cultural Discovery (4 Days)**
* **Focus:** Temples, gardens, traditional experiences, and culinary exploration.
* **Day 1: Arrival & Gion Charm.**
* Afternoon: Check into accommodation, explore Gion district (geisha spotting).
* Evening: Traditional Kaiseki dinner.
* **Day 2: Iconic Temples & Bamboo Forest.**
* Morning: Kinkaku-ji (Golden Pavilion).
* Lunch: Local udon restaurant.
* Afternoon: Arashiyama Bamboo Grove & Tenryu-ji Temple.
* Evening: Relaxed dinner near accommodation.
* **Day 3: Fushimi Inari & Nishiki Market.**
* Morning: Fushimi Inari Shrine (hike through torii gates).
* Lunch: Explore Nishiki Market for diverse street food.
* Afternoon: Kiyomizu-dera Temple for panoramic views.
* Evening: Optional tea ceremony experience.
* **Day 4: Departure.**
* Morning: Last-minute souvenir shopping or revisit a favorite spot.
****
**USER'S TRAVEL GOAL:**
**Destination:** Paris, France
**Duration:** 4 full days
**Travelers:** 2 adults, 1 child (age 12)
**Interests:** Iconic landmarks (Eiffel Tower, Louvre), engaging activities for the child, delicious pastries and French cuisine, relaxed pace, avoiding long queues where possible.
**Accommodation:** Centrally located hotel near public transport.
**YOUR TASK:**
Create a detailed, day-by-day itinerary for this family trip to Paris. For each day:
1. **Morning Activity:** Specify the main attraction or activity.
2. **Lunch Suggestion:** Recommend a type of eatery or specific area known for good, family-friendly food near the morning activity.
3. **Afternoon Activity:** Detail a secondary activity or exploration.
4. **Evening Plan:** Suggest dinner options and any light evening entertainment.
5. **Checkpoint/Tip:** Include a practical tip for each day, focusing on queue management, child engagement, or local flavor.
Ensure the itinerary balances famous sights with experiences suitable for a 12-year-old, incorporates opportunities for enjoying French cuisine and pastries, and maintains a relaxed pace.
Enhancing the Preference Extractor in Multi-turn Dialogues
Enhancing the Preference Extractor in Multi-turn Dialogues: From Annotating Disasters to Accurate Preference Extraction, arXiv:2508.01739
Method of iterative preference tracking in dialogues. Instead of passing the entire chat history, the model receives a brief summary of established preferences and only the last user message. The task is simplified to incremental state updates, which prevents loss of critical details and contradictions in long conversations.
Use it whenUse the IterChat method when developing dialogue systems with long multi-turn conversations where it is critical to preserve details of user preferences. Especially effective for trip planning, personal recommendations, consulting services, where the user gradually clarifies requirements. Apply when you notice that the LLM forgets previously mentioned details (budget, dates, preferences) or contradicts itself in long dialogues. Suitable for chatbots, personal ass
Prompt
You are an expert travel planning assistant. Your task is to help a client plan a trip to Italy by iteratively refining their preferences and plan. You will manage the conversation by maintaining a clear "State" of their preferences and updating it based on their "Current Request".
**State Update Rules:**
1. **Preserve Unchanged Preferences:** Always retain all information from the `[Previous State]` that is not explicitly modified or removed by the `[Current Request]`.
2. **Incorporate New Information:** Add any new preferences or details mentioned in the `[Current Request]`. If a preference already exists, update it with the new information.
3. **Handle Modifications/Removals:** If the user explicitly states to "forget about," "remove," or "change" a preference, update the state accordingly.
4. **Output Format:** After processing the `[Current Request]`, output the complete, updated `[Current State]` in JSON format.
---
**DIALOGUE CONTEXT:**
**[Previous State]:**
{
"Destination": "Italy",
"Duration": "10 days",
"Budget per person": "1500 euros",
"Interests": ["history", "food"]
}
**[Current Request]:**
"Thanks for the initial plan! I've been thinking, and I'd really like to include a visit to a famous beach and perhaps try some scuba diving. Also, let's increase the budget to 1800 euros per person to allow for more flexibility."
---
**Your Task:**
Analyze the `[Current Request]` and update the `[Previous State]`. Output the final, updated state as a JSON object.
Exploring ReAct Prompting for Task-Oriented Dialogue: Insights and Shortcomings, arXiv:2412.01262
ReAct forces LLM to follow a "Thought → Action → Observation" cycle, turning the model into a systematic planner. Instead of answering directly, the model first formulates a plan, performs an action (API request, database query), analyzes the result, and only then responds to the user, which improves dialogue quality.
Use it whenUse ReAct when developing conversational agents or chatbots for tasks requiring interaction with external systems (databases, APIs). Particularly effective for booking scenarios, consultations, customer support, where both accuracy and communication quality matter. Apply when user experience and natural dialogue are more important than absolute formal accuracy. Suitable for cases requiring step-by-step action planning with the ability to receive intermediate results.
TripTailor: A Real-World Benchmark for Personalized Travel Planning, arXiv:2508.01432
Decomposition of a complex planning task into a sequence of simple steps: gathering options, ranking, assembly, optimization. Each prompt solves one specific subtask, which allows the LLM to better account for multiple constraints and create realistic, executable plans instead of beautiful but useless fantasy.
Use it whenApply this technique when an LLM needs to solve a complex task with multiple constraints and conditions: travel planning, event organization, project plan creation, resource allocation. Especially effective when a one-step prompt produces unrealistic results, ignores part of the requirements, or creates logically contradictory plans. The method is critical for tasks requiring consideration of budget, time, geography, personal preferences, and other interconnected parameters simultan
LISTEN to Your Preferences: An LLM Framework for Multi-Objective Selection, arXiv:2510.25799
Framework for decision-making with multiple choices. Structures the prompt by 5 components: role, criteria, priorities, options, task. LLM weighs trade-offs between criteria based on informal description of preferences, without mathematical formulas. Iteratively selects the best option through weight refinement or tournament comparison.
Use it whenUse LISTEN when you need to choose one option from many alternatives (10-100+ options) by several competing criteria. Ideal for situations where it's difficult to formalize preferences mathematically: supplier selection, candidate selection, trip planning, equipment procurement. Especially effective when criteria conflict (price vs quality, speed vs reliability) and priorities are expressed informally. Suitable for tasks where transparency of reasoning and consideration of subjective preferences are important
TripTide: A Benchmark for Adaptive Travel Planning under Disruptions, arXiv:2510.21329
A framework of structured prompting for adapting plans when failures occur. Uses a three-part structure: full context of the original plan, classification of failure scale, and response profile. Allows obtaining relevant solutions instead of random fixes, precisely matching constraints and priorities.
Use it whenUse TripTide Framework when you need to adapt an existing plan when unforeseen circumstances or failures occur. Particularly effective for project planning, event organization, content planning, budgeting, and any scenarios requiring meaningful strategy adjustment. Apply when simple requests like 'fix the plan' give superficial results and precise adaptation is needed considering constraints, priorities, and scale of changes. Suitable for situations of various.
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.