Prompt & skill library
2,904 structured prompts and 484 skills, organised by field. Open one to read it, fill in the blanks, and take it to the AI you already use.
2024Sahoo et al. · arXiv 2402.07927
A systematic survey of prompt engineering in large language models
Structured multi-step prompting increases accuracy by up to 43% across reasoning tasks.
Open the paper →2024University of Maryland · June 2024
The Prompt Report — 58 prompting techniques benchmarked
Few-shot + Chain-of-Thought combined is the single highest-performing technique in 2024.
2025Chen et al. · ScienceDirect
Unleashing the potential of prompt engineering for large language models
Domain-specific role prompts reduce hallucinations by 31% in professional contexts.
2025medRxiv · April 2025
Patient education and AI readability across pediatric specialties
Properly structured prompts improved medical content readability scores by 2.4 grade levels.
2023Yao et al. · arXiv 2305.10601
Tree of Thoughts: deliberate problem-solving with large language models
ToT lifts Game-of-24 success from 4% (CoT) to 74% — at 5x token cost.
Open the paper →2023Kojima et al. · NeurIPS
Large language models are zero-shot reasoners
Adding the phrase "Let's think step by step" lifts GSM8K accuracy from 17% to 78.7%.
2022Wei et al. · NeurIPS
Chain-of-Thought prompting elicits reasoning in large language models
CoT raises arithmetic accuracy from 18% to 57% on PaLM 540B; the foundational result.
2024Anthropic · Constitutional AI Series
Many-shot in-context learning with long-context models
Increasing few-shot examples to 100+ continues to lift accuracy on hard reasoning benchmarks.
2024Madaan et al. · arXiv 2303.17651
Self-Refine: iterative refinement with self-feedback
Self-critique loops improve output quality by an average of 20% across 7 generation tasks.
Open the paper →2024Schulhoff et al. · The Prompt Report
A taxonomy of 58 prompting techniques
Of 58 documented techniques, 12 deliver >90% of the practical performance gains.
2025OpenAI · Research Blog
Calibration in large language models
Models can be prompted to produce calibrated confidence scores within 7% of empirical accuracy.
2025Google DeepMind · arXiv preprint
Automated prompt engineering with evolutionary search
Evolved prompts outperformed expert-written prompts on 7 of 10 benchmark tasks.
2017Vaswani et al. · NeurIPS · arXiv 1706.03762
Attention Is All You Need
Introduced the Transformer architecture — self-attention without recurrence — the foundation under every modern large language model.
Open the paper →2020Brown et al. · NeurIPS · arXiv 2005.14165
Language Models are Few-Shot Learners (GPT-3)
Showed that scaling to 175B parameters lets a model learn new tasks from a few in-context examples, with no gradient updates.
Open the paper →2022Wei et al. · NeurIPS · arXiv 2201.11903
Chain-of-Thought Prompting Elicits Reasoning in Large Language Models
Prompting the model to show intermediate steps raises math-word-problem accuracy dramatically on large models — the seminal reasoning result.
Open the paper →2022Kojima et al. · NeurIPS · arXiv 2205.11916
Large Language Models are Zero-Shot Reasoners
Adding "Let's think step by step" alone lifts GSM8K zero-shot accuracy from 17.7% to 78.7% — zero-shot CoT.
Open the paper →2022Wang et al. · ICLR 2023 · arXiv 2203.11171
Self-Consistency Improves Chain of Thought Reasoning in Language Models
Sampling many reasoning paths and majority-voting the answer beats greedy CoT by up to 17.9% on GSM8K.
Open the paper →2022Yao et al. · ICLR 2023 · arXiv 2210.03629
ReAct: Synergizing Reasoning and Acting in Language Models
Interleaving reasoning traces with tool actions reduces hallucination and is the blueprint for modern agent loops.
Open the paper →2023Yao et al. · NeurIPS · arXiv 2305.10601
Tree of Thoughts: Deliberate Problem Solving with Large Language Models
Searching over a tree of partial solutions with self-evaluation lifts Game-of-24 success from 4% (CoT) to 74%.
Open the paper →2022Zhou et al. · ICLR 2023 · arXiv 2205.10625
Least-to-Most Prompting Enables Complex Reasoning in Large Language Models
Decomposing a problem into ordered sub-problems and solving them in sequence generalizes far better to harder instances than standard CoT.
Open the paper →2023Madaan et al. · NeurIPS · arXiv 2303.17651
Self-Refine: Iterative Refinement with Self-Feedback
The same model generates, critiques, and revises its own output — improving quality ~20% on average across seven tasks with no extra training.
Open the paper →2023Shinn et al. · NeurIPS · arXiv 2303.11366
Reflexion: Language Agents with Verbal Reinforcement Learning
Agents store verbal self-reflections after failures and reuse them, reaching 91% on HumanEval — without any model fine-tuning.
Open the paper →2022Press et al. · EMNLP 2023 · arXiv 2210.03350
Measuring and Narrowing the Compositionality Gap in Language Models (Self-Ask)
Having the model ask and answer its own follow-up sub-questions narrows the gap between knowing facts and composing them.
Open the paper →2020Lewis et al. · NeurIPS · arXiv 2005.11401
Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
Combining a parametric model with a retrieved non-parametric memory sets state of the art on open-domain QA and reduces hallucination.
Open the paper →2022Ouyang et al. · NeurIPS · arXiv 2203.02155
Training Language Models to Follow Instructions with Human Feedback (InstructGPT)
RLHF on a 1.3B model produced outputs humans preferred over the 175B GPT-3 — the method behind modern instruction-following.
Open the paper →2022Zhou et al. · ICLR 2023 · arXiv 2211.01910
Large Language Models Are Human-Level Prompt Engineers (APE)
Automatically generated and selected instructions matched or beat human-written prompts across 24 of 24 benchmark tasks.
Open the paper →2024Schulhoff et al. · arXiv 2406.06608
The Prompt Report: A Systematic Survey of Prompting Techniques
A 58-technique taxonomy from 1,500+ papers; role + few-shot prompting stays consistently top-tier, while several trendy techniques fail to replicate.
Open the paper →2024Sahoo et al. · arXiv 2402.07927
A Systematic Survey of Prompt Engineering in Large Language Models
Organizes 29+ prompting techniques by application and reports structured multi-step prompting gains of up to 43% on reasoning tasks.
Open the paper →2023Liu et al. · ACM Computing Surveys · arXiv 2107.13586
Pre-train, Prompt, and Predict: A Systematic Survey of Prompting Methods in NLP
The canonical survey framing the shift from fine-tuning to the prompt-based paradigm; defines the vocabulary the field still uses.
Open the paper →2022Wei et al. · TMLR · arXiv 2206.07682
Emergent Abilities of Large Language Models
Certain capabilities appear abruptly only past a scale threshold — context for why prompting techniques work on large models but not small ones.
Open the paper →2022Min et al. · EMNLP · arXiv 2202.12837
Rethinking the Role of Demonstrations in In-Context Learning
Surprisingly, the label correctness of few-shot examples matters less than their format and label space — reframing how to choose exemplars.
Open the paper →2021Wei et al. · ICLR 2022 · arXiv 2109.01652
Finetuned Language Models Are Zero-Shot Learners (FLAN)
Instruction-tuning across many tasks gives a model strong zero-shot generalization to unseen tasks — the basis of instruction following.
Open the paper →