00 Start

Prompt tips are everywhere. Receipts are not

This guide only includes techniques with a named author and a source you can check.

The internet is drowning in "10 secret prompts" posts with zero evidence. Meanwhile, the techniques that actually move results have papers, authors and measurements behind them — most people just never see the receipts.

So here's the deal: every technique below comes with who invented or measured it, when, and a link to the primary source. And every technique comes with a try-it prompt filled with your own task — because reading about prompting changes nothing.

What you'll walk away with

• 8 techniques with receipts: few-shot, chain-of-thought (and when it hurts), self-consistency, grounding, step-back, roles, chaining, context engineering
• Try-it prompts pre-filled with your task — run them as you go
• The map: where to find the other 50 techniques when you're ready

01 Setup

Pick one real task — your test bench

Every technique below gets tried on the same task, so you can feel the differences.

Task: [TASK]. Field: [FIELD].

02 The classics

Few-shot prompting — the oldest receipt in the book

Source: Brown et al., "Language Models are Few-Shot Learners" (OpenAI, 2020) — the GPT-3 paper.

The founding discovery of modern prompting: models learn from examples inside the prompt, no retraining needed. Showing 2–5 examples of the output you want beats describing it in adjectives — still true on every current model. Original paper (arXiv:2005.14165).

Do this task: [TASK] — in the same style, structure and level of specificity as these examples:

[paste 2–4 examples of output you love — other niches welcome]

Keep what makes the examples strong, but don't copy their wording.

02 The classics

Chain-of-thought — and the 2026 plot twist

Sources: Wei et al. (Google, 2022) and Kojima et al. (2022). Plus a warning label.

The receipt: Wei et al. showed that demonstrating step-by-step reasoning dramatically improves complex-task accuracy (arXiv:2201.11903). Kojima et al. then found the magic five words — "Let's think step by step" — worked with zero examples (arXiv:2205.11916).

The plot twist: on modern reasoning models (which think before answering by design), explicitly commanding step-by-step thinking is redundant — and vendor docs now warn it can even hurt. What survives in 2026 is the deeper principle: decompose the task, don't micromanage the thinking.

Before doing this task: [TASK] — first analyze what makes it hard in my field ([FIELD]): the audience, the constraints, the failure modes. Show the analysis briefly, then do the task based on it.

03 Reliability

Self-consistency — ask three times, trust the overlap

Source: Wang et al. (Google, 2022) — arXiv:2203.11171.

The finding: sampling several independent reasoning paths and taking the majority answer beats any single answer (original paper). The DIY version costs you one prompt:

Do this task three times, independently, taking a different angle each time: [TASK].

Then compare the three results: what do they agree on (that's likely solid), where do they diverge (that's where the risk lives) — and assemble a final version that keeps the agreements and resolves the divergences explicitly.

03 Reliability

Grounding: "according to..." — three words against hallucinations

Source: Weller et al. (2023) — arXiv:2305.13252.

The measured effect: anchoring a request to a named source ("According to Wikipedia…") makes the model quote real grounded content more and invent less (original paper). Combine it with an honesty exit and you get a practical anti-hallucination kit:

Answer based on the attached materials / on [named source] only: [TASK].

Rules: if the source doesn't cover something — say "not in the source" instead of filling the gap. At the end, list which claims come from the source and which are your inference.

04 Perspective

Step-back prompting — zoom out before you zoom in

Source: Zheng et al. (Google DeepMind, 2023) — arXiv:2310.06117.

The finding: asking the model to first answer a more general "step-back" question — the principle behind the task — and only then the task itself improves reasoning accuracy (original paper).

Before doing my task, answer a step-back question first: what are the general principles of doing this well in [FIELD]? List 3–5 principles.

Now, applying those principles, do the task: [TASK] — and note which principle each key decision comes from.

04 Perspective

Role prompting — the technique with an asterisk

Source: The Prompt Report — Schulhoff et al. (2024), the 58-technique survey.

Honest receipt: the most comprehensive survey of prompting (1,565 papers reviewed) found that role prompting has mixed evidence for accuracy gains — "you are an expert" doesn't reliably make answers smarter (arXiv:2406.06608).

Where roles DO earn their keep: perspective and voice. A role changes what the model pays attention to — which is exactly what you want for critique:

You are my toughest customer in [FIELD] — skeptical, busy, allergic to marketing. Look at the result of my task below and tell me honestly: what works, what triggers doubt, what's missing, what you'd ignore completely.

Result: [paste the output you got from an earlier technique]

05 The shift

Prompt chaining — outputs feeding inputs

Source: Anthropic's prompt engineering docs.

Vendor-documented and battle-tested: complex work splits into a chain where each step's output becomes the next step's input — research → draft → critique → revision — instead of one mega-prompt trying to do everything (Anthropic's best practices).

Link 1: "Analyze the audience and constraints for: [TASK]"
Link 2: "Based on that analysis, do the task"
Link 3: "Now critique the result as my toughest customer"
Link 4: "Revise, addressing the critique"

(Run as four messages in one chat — each link reads the previous output.)

05 The shift

Context engineering — where the field moved in 2025–2026

Coined into the mainstream by Tobi Lütke (Shopify CEO) and Andrej Karpathy, mid-2025.

Karpathy's analogy became the dominant mental model: "The LLM is a CPU, the context window is RAM, and you're the operating system." The wording of the request matters less than what you load into the window: your materials, examples, constraints, standards. Anthropic, LangChain and LlamaIndex formally adopted the framing within weeks (Context Engineering Guide 2026).

Same task, third run: [TASK].

This time I'm changing nothing in the wording — only adding context: [attach 2–3 of your best past outputs, your audience notes, your "never do this" list].

Use these materials as the primary guide. Compare in your head: which moved the result more — the fancy wording, or the materials?

06 Finale

Done. 8 techniques, 8 receipts, 1 task that felt them all

The full map — when you're ready for more

• The Prompt Report — Schulhoff et al., a systematic survey of 58 text prompting techniques: arXiv:2406.06608 · readable summary at learnprompting.org
• Anthropic's prompting best practices: platform.claude.com docs — including the model-specific Fable 5.1 guide
• Prompting the newest model specifically — the Claude Fable 5.1 guide