Prompt Engineering Toolkit

# Prompt Engineering Toolkit

A modular, evidence-based template system for writing effective LLM prompts.

---

## 1. Design Principles

Each principle below is grounded in findings from cross-provider documentation (OpenAI, Anthropic, Google), peer-reviewed research, and validated practitioner patterns.

### P1 — Task Specificity Over Vagueness

Describing **what the output should look like** beats describing **how good it should be**. "Clear and under 100 words" outperforms "professional and polished" every time. Bsharat et al. (2023) measured a 57.7% quality improvement when explicit format constraints were added — even against deliberately weak baselines, the direction is unambiguous. All three major providers independently recommend this as the single highest-ROI prompting habit.

### P2 — Structure Before Length

Structured short prompts consistently outperform verbose unstructured ones. Google's Gemini 3 guide warns the model "may over-analyze verbose or overly complex prompt engineering techniques." OpenAI's GPT-5.4 guidance: "Start with the smallest prompt that passes your evals." The template system below is therefore modular — start minimal, add blocks only when they solve a measured problem.

### P3 — Information Placement Matters

The "Lost in the Middle" finding (Liu et al. 2023, 2500+ citations) shows a U-shaped attention curve: models attend far more to content at the **beginning and end** of prompts. Place instructions and constraints at the start. Place input data after instructions. Place your most important constraint or quality bar last if it's short.

### P4 — Positive Framing

All three providers recommend telling the model **what to do**, not what to avoid. Negative instructions force the model to process the unwanted concept, potentially increasing its probability — Anthropic calls this the "Pink Elephant Problem." Google warns broad negatives like "do not infer" can break basic reasoning.

**Fix:** "Write in plain language accessible to a general audience" instead of "Don't use jargon."

### P5 — Examples Beat Descriptions

Few-shot examples (2–5) are among the highest-ROI techniques. Min et al. (2022) found that format, distribution, and label space of examples matter more than whether individual labels are correct. But more is not always better — the "over-prompting" effect (2025) shows performance peaks then declines with excessive examples.

**Rule of thumb:** Start with 2–3 high-quality diverse examples. Add more only if output quality improves.

### P6 — Decompose Complex Tasks

Breaking a complex prompt into sub-tasks — either as numbered steps within one prompt or as a chain of separate prompts — is supported by all three providers. OpenAI measured +20% agentic pass rate when planning instructions were added. Academic decomposed-prompting research confirms the mechanism.

### P7 — Permit Uncertainty

Instructing the model to say "I don't know" or "I'm not sure" when uncertain reduces hallucination across all tested models. This is cheap, effective, and under-used. All three providers explicitly recommend it.

### P8 — Match Delimiters to Model

Claude was trained on XML tags. GPT prefers Markdown headers. Gemini accepts both but demands consistency. Using the right format for your model doesn't change whether the prompt works, but it affects how well it works. The templates below use model-agnostic section headers that can be trivially converted.

### P9 — Iterate Empirically

Every provider describes prompt engineering as "inherently empirical." The best practitioners run experiments and take notes. A prompt that works on GPT-4o may degrade on GPT-5. The template system below includes a review step not because it's always needed, but because it's cheap insurance for high-stakes tasks.

---

## 2. Master Multi-Purpose Prompt Template