# Inside the Prompt Engineering Guide repo: Nextra, two lockfiles, and no version to pin > DAIR.AI's Prompt Engineering Guide is an MIT-licensed documentation website rather than an installable library. It maps named prompting techniques, RAG patterns and agent material, but the repository ships no evaluation harness, no releases and a package manifest still named after someone else's template. **dair-ai/Prompt-Engineering-Guide** — GitHub describes it as 🐙 Guides, papers, lessons, notebooks and resources for prompt engineering, context engineering, RAG, and AI Agents.. The repository metadata lists MDX as its primary language. The metadata lists the MIT license. This article stays within the project description and details documented in the GitHub repository README. - Repository: https://github.com/dair-ai/Prompt-Engineering-Guide - Website: https://www.promptingguide.ai/ - Stars: 78,721 · Forks: 8,652 - Language: MDX - License: MIT - Published: 2026-08-13 · Updated: 2026-08-18 · Language: en - Canonical page: https://hysenlabs.com/projects/dair-ai-prompt-engineering-guide ## The site is a Nextra build still carrying the template's name This repository is MDX, not an importable library, and nothing in it can be added as a dependency. What ships is a documentation website. The manifest at the root makes the origin plain: package.json names the project nextra-docs-template, gives it the description Nextra docs template, points its repository field at git+https://github.com/shuding/nextra-docs-template.git, and credits Shu Ding as author. Those four lines are template metadata that survived the fork into DAIR.AI. The runtime dependencies confirm what the thing is. Next is pinned at ^13.5.6, nextra and nextra-theme-docs at ^2.13.2, React at ^18.2.0, with katex at ^0.16.27 for the mathematics in the technique pages, clsx for class names, FontAwesome icons, and @vercel/analytics. The only three scripts are dev, build and start, all of which wrap a Next command. A reader who expects a prompt library to install will find a docs site with a build step. The stale template name has a practical cost for anyone who vendors the repository. The package identifier written into a lockfile, a container image name or a build log is nextra-docs-template, a string that points a search at an unrelated project. Worse, the repository URL in the manifest is a third party's, so automated tooling that derives provenance from package.json will record the wrong owner. ## Two lockfiles sit at the root with nothing to break the tie The root contains package-lock.json and pnpm-lock.yaml at the same time. package.json has no packageManager field, and none of the three scripts declares which installer is expected. Two contributors working from the same commit with different package managers resolve next ^13.5.6 and nextra ^2.13.2 through different algorithms and can finish with different dependency trees. For most readers this is trivia, because the deployed site at promptingguide.ai is the product and nobody installs the repository to read it. For a team mirroring the guide internally, or for a fork that adds pages, the absence is the first real decision the repository refuses to make for you. Pin the manager yourself in the fork rather than inheriting the ambiguity. ## Content is split across pages/, guides/ and a directory named ar-pages/ The top level mixes routing, content and assets in a layout the README never explains. Alongside pages/ and guides/ there is a separate directory named ar-pages/, plus components/, public/, img/, notebooks/ and lecture/. The README claims support for 13 languages, and the presence of a parallel directory with a language prefix in its name is the physical trace of that claim: the English material and the non-English material are separate trees, not one tree with translated front matter. That split is the first thing to check before editing a page. Nothing in the repository joins the trees. There is no translation script, no sync command among the three npm scripts, and no note in the README describing how a language is added or kept aligned. A correction applied under pages/ or guides/ therefore has no documented route to its counterpart elsewhere, and the languages can drift apart without anything failing loudly. Two further root files point the same way: CITATION.cff, the machine-readable record for citing the guide in academic work, and CLAUDE.md, which appears in the tree and is never mentioned in the README, so a contributor has no way to tell what instructions it carries. ## Seventeen technique names and no order to try them in The techniques index is the guide's centre of gravity, and it is a list of names. Zero-Shot, Few-Shot, Chain-of-Thought, Self-Consistency, Generate Knowledge, Prompt Chaining, Tree of Thoughts, Retrieval Augmented Generation, Automatic Reasoning and Tool-use, Automatic Prompt Engineer, Active-Prompt, Directional Stimulus Prompting, Program-Aided Language Models, ReAct, Multimodal CoT and Graph Prompting each get their own page, and the applications index adds Function Calling, Generating Data, Generating Synthetic Dataset for RAG and Generating Code alongside a Graduate Job Classification case study. What the index never supplies is a decision procedure. No entry says which technique to reach for first, what any of them costs in tokens or round trips, or which ones compose with each other. A reader arriving with a concrete problem has to open pages one at a time to discover the fit, and each page is the only place that judgement exists. The Prompt Hub is indexed the same way, by task rather than by difficulty, with entries for Classification, Coding, Creativity, Evaluation, Information Extraction, Image Generation, Mathematics, Question Answering, Reasoning, Text Summarization, Truthfulness and Adversarial Prompting. Grouping by task is the right cut for someone who already knows the technique and wants a template. For someone who does not, it hides the prerequisite. ## The top third of the README is a storefront Before the index there is a SerpApi sponsorship line, a link to paid self-paced courses at academy.dair.ai with the discount code PROMPTING20 for 20 percent off, and a pointer to corporate training, consulting and talks at promptingguide.ai/services. The guide itself is free and MIT licensed, but the surrounding page is a sales funnel and the two live in the same file. Two numbers in that block deserve scepticism from an engineer deciding whether to depend on this. The README says the project crossed 3 million learners in January 2024, and that it reached number one on Hacker News on 21 Feb 2023. Both are self-reported, neither cites a source, and neither defines what a learner is. The milestone tells you the guide was widely read at some point. It says nothing about whether the technique pages are current, and the site publishes no per-page dates to check them against. ## No releases, so there is no version of the guide to pin The repository has no GitHub releases at all, and package.json carries the version 0.0.1 that came from the template. The default branch is main, and the last push to it was on 2026-03-11. That combination leaves the content addressable only as a moving commit. There is no tag, no changelog, and no release note that would tell you a technique page was rewritten, a technique was removed, or a paper link went dead. If your team cites the guide in an internal playbook, cite a commit hash and a date rather than a version string, because 0.0.1 carries no information about content. The same argument applies to any page you rely on: copy the text you need into your own notes, since a page can be edited in place on main without a signal reaching you. ## Nothing in the repository measures whether a prompt change worked This gap decides who should read the guide and who should skip it. The dependency list is a documentation toolchain and nothing else: Next, Nextra, React, KaTeX, clsx, FontAwesome, Vercel analytics. There is no evaluation harness, no prompt regression suite, no scoring script, and no model SDK among the packages. Directories named notebooks/ and lecture/ hint at experiments and material that once lived here, but the top level contains nothing that runs an experiment against a model. So the guide can tell you that a technique exists and roughly what shape it takes. It cannot tell you whether the technique helped your workload, and nothing in the repository will notice when a provider changes model behaviour underneath a prompt you copied from a page. Anyone shipping prompts to production needs a measurement loop of their own; treating a cited page as evidence of quality means trusting a list of references with no numbers attached. The closest thing to an interactive artefact is a standalone file named infographic-review.html at the repository root, whose purpose the README does not describe and whose relationship to the site you would have to infer from the name. ## Conclusion Treat the Prompt Engineering Guide as an organised map of named techniques rather than a dependency: there is nothing to install, nothing to import, and no version to pin. It suits an engineer who needs vocabulary and a starting point for RAG, function calling or chain-of-thought design. It does not suit anyone who needs measured proof that a prompt change improved a product, because the repository contains no evaluation harness and the last push to main was on 2026-03-11. Before you build a curriculum or an internal reference page on it, open the specific technique page you depend on and confirm it still says what you remember, then check whether the technique list has gained or lost an entry since you last read it. ## FAQ ### What are the 5 principles of prompt engineering? This repository does not enumerate a five-item principle list. Its introductory advice is split across separate pages named Basics of Prompting, Prompt Elements and General Tips for Designing Prompts, so there is no single page here that presents five principles under that heading. ### How should I learn prompt engineering? Start with the web version at promptingguide.ai, which is where the README sends readers for the current material. The introduction track runs LLM Settings, Basics of Prompting, Prompt Elements, General Tips for Designing Prompts and Examples of Prompts before the techniques pages begin. Paid self-paced courses sit separately at academy.dair.ai. ### What are the 5 P's of effective prompting? Nothing in this repository is organised under that name. The closest structure is the Prompt Hub, which groups ready-made prompts by task, with sections for Classification, Coding, Creativity, Evaluation, Information Extraction, Mathematics, Question Answering, Reasoning, Text Summarization, Truthfulness and Adversarial Prompting, rather than by a set of P-named principles. ### Is there a PDF version of the ChatGPT prompt guide available? The guide is distributed as a website at promptingguide.ai, and its Models section includes a ChatGPT page at promptingguide.ai/models/chatgpt. The repository has no PDF export step: the only scripts in package.json are dev, build and start, and each one runs a Next command rather than generating a document. ## Sources - [Official documentation](https://www.promptingguide.ai/) - [Official README](https://github.com/dair-ai/Prompt-Engineering-Guide#readme) - [Project repository](https://github.com/dair-ai/Prompt-Engineering-Guide) --- Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/dair-ai-prompt-engineering-guide