Prompt Engineering Guide (DAIR.AI)

dair-ai‘s open reference for prompting technique, and the largest source in this spoke after developer-roadmap: 77.4k stars, 8.5k forks, MIT, 1,589 commits on main as of 2026-08-09. It describes itself as “guides, papers, lessons, notebooks and resources for prompt engineering, context engineering, RAG, and AI Agents,” ships as a repository plus a site at promptingguide.ai, and claims “We crossed 3 million learners in January 2024.” Thirteen languages. Parked in the hub _inbox on 2026-07-26; routed here 2026-08-09, the boundary case this spoke’s own registry note left open.

What is in it

Introduction — LLM settings, basics of prompting, prompt elements, design tips, example prompts.
Techniques — 16, in a flat list: zero-shot, few-shot, chain-of-thought, self-consistency, generate-knowledge, prompt chaining, tree of thoughts, RAG, ART, APE, Active-Prompt, DSP, PAL, ReAct, multimodal CoT, graph prompting.
Applications — function calling, data generation, synthetic datasets for RAG, code generation, a job-classification case study.
Prompt Hub — ready prompts filed under 12 use cases (classification, coding, creativity, evaluation, information extraction, image generation, mathematics, question answering, reasoning, summarization, truthfulness, adversarial prompting).
Model pages — ChatGPT, Code Llama, Flan, Gemini, GPT-4, LLaMA, Mistral 7B, Mixtral, OLMo, Phi-2, plus a model collection.
Risks — adversarial prompting, factuality, biases.

Delivery is markdown lessons and a website, with Jupyter notebooks and a one-hour recorded lecture beside them. There is a Discord, a newsletter, and a DAIR.AI Academy selling self-paced courses, corporate training, consulting and talks.

A third form: written to be consulted, not completed

The spoke has two forms so far — the meta-curriculum that points at other people’s material, and the executable-curriculum whose lessons are code you run. This is neither. The lessons are DAIR.AI’s own writing, so it is not curation; nothing has to be executed to read it, so it is not construction.

What it is, structurally, is a reference. The Prompt Hub is indexed by the task you arrived with. The model pages are indexed by the model you happen to be using. The 16 techniques sit in a flat list with no prerequisites and no order, and nothing in the repository records where a reader is or what they have finished. That is the design of a handbook someone opens at the page they need, and it is worth separating from the sequenced curricula around it, because the two are not doing the same job. It also makes the headline number soft in a specific way: “3 million learners” counts arrivals, and a reference gets most of its arrivals from people looking one thing up.

The controlled comparison this spoke wanted

nirdiamant-prompt-engineering is the same subject — prompting technique — with the same absence of an exam behind it, and it made the opposite pedagogical choice: 22 runnable notebooks against this one’s written explanation and lookup tables. Two curricula, same domain, same missing assessment regime, different form.

The synthesis had narrowed its founding correlation to where an assessment regime exists it shapes scope; pedagogy is chosen on other grounds. This pair is the first evidence in the corpus that is actually controlled rather than merely consistent with it. Holding the subject fixed and the exam fixed, the pedagogy still varies — so whatever picks the form, it is not the assessment regime. The two also sit 10× apart in stars (77.4k against 7.8k), which says something about what readers want from a technique domain and nothing about which one teaches better.

Two things to hold against it

The commercial layer is now the rule, and this extends it past individuals. Every self-published source in this spoke has something to sell beside the free artifact. Here the seller is an organization with an Academy, corporate training and a consulting line, which is a larger apparatus than nir-diamant‘s book or kamran-ahmed‘s site. The free guide is the top of a funnel. That does not make the content worse; it does mean the guide’s own account of its reach is marketing copy.

The reference form decays unevenly. A meta-curriculum rots by link rot and an executable one by dependency drift. This one has two clocks: the technique sections age slowly — chain-of-thought and ReAct are still chain-of-thought and ReAct — while the model pages age fast, and the ones on offer here (GPT-4, Phi-2, Mixtral, Flan, OLMo) are a 2023–2024 shelf being consulted in 2026. The “3 million learners” line has sat unchanged since January 2024 in the same way. The parts most likely to be looked up in a hurry are the parts most likely to be stale.

Provenance note

T3. Tiered by what its claims rest on rather than by who published them: the technique write-ups summarize published papers and are checkable, but the guide’s claims about itself — reach, learner counts, effectiveness — are self-reported by a party selling courses on the strength of them. No outcome data of any kind, which is the spoke’s standing hole rather than this source’s failing.