Spokes.wiki Search About
Software Source Code source ↗ source url updated Wed Aug 05 2026 00:00:00 GMT+0000 (Coordinated Universal Time)

Prompt Engineering (NirDiamant)

nir-diamant‘s course in prompt engineering: 22 Jupyter notebooks, each one technique, all runnable against the OpenAI API through LangChain. 7,752 stars and 1,006 forks as of 2026-08-05; created 2024-10-10 and last pushed 2026-07-31, so it is maintained rather than abandoned. Routed here from the hub on 2026-08-05.

The sequence

The README groups the 22 into seven tiers of stated difficulty: Fundamental Concepts (introduction, basic structures, templates with Jinja2) → Core Techniques (zero-shot, few-shot, chain-of-thought) → Advanced Strategies (self-consistency, constrained generation, role prompting) → Advanced Implementations (task decomposition, prompt chaining, instruction engineering) → Optimization and Refinement (optimization, ambiguity, length and complexity) → Specialized Applications (negative prompting, formatting, task-specific prompts) → Advanced Applications (multilingual, ethics, security, evaluation).

The sequence exists only in the README. All 22 notebooks sit in one flat all_prompt_engineering_techniques/ directory with no numeric prefixes, so nothing in the repository enforces the order or records where a reader is. Compare ai-engineering-from-scratch, whose 20 phases are structural. The ordering here is a recommendation.

It is neither of this spoke’s two known forms

Not a meta-curriculum — the lessons are the author’s own, not links to other people’s. Not build-from-scratch either: a reader runs ChatOpenAI and Jinja2 templates, not an implementation of anything. The lesson is executable code you edit and re-run, a third form the spoke had not seen before (executable-curriculum).

Why it matters to the founding thesis

The spoke’s synthesis rests on one correlation across three sources: curricula with an exam behind them teach by curation, and the one without an exam teaches by construction. This source is a fourth data point and it falls on the predicted side — there is no prompt-engineering interview to prepare for, and the pedagogy is constructive.

That is weaker support than it first looks. Prompt engineering is a technique domain, not an engineering discipline with a hiring pipeline, so the absence of an exam is close to guaranteed by the subject rather than chosen by the author. The correlation still wants a case that could have gone either way: a curated-pointer curriculum with no assessment behind it, or a build-it one aimed at an exam.

Licensing and money

A custom non-commercial license (GitHub reads it as NOASSERTION) — the most restrictive terms in the corpus, against GPL-3.0 for awesome-system-design-resources. Beside the repository sit a published book, a newsletter the README puts above 50,000 subscribers, a YouTube channel, and Discord and Reddit communities. Diamant also maintains sibling repositories on RAG techniques, GenAI agents and agent memory.

Three of this spoke’s four sources now have a commercial product adjacent to the free curriculum. That is worth recording as the normal condition of the genre rather than as a mark against any one of them.

Where the subject stops being this spoke’s business

No spoke owns prompt engineering as a discipline. The hub parked DAIR.AI’s Prompt Engineering Guide on 2026-07-26 for exactly that reason and it is still in _inbox; this source is the second prompting arrival. It routed here because it is shaped as a course — ordered, tiered, hands-on — where the DAIR.AI guide is shaped as a handbook, indexed by technique for lookup.

So the claims inside the notebooks — whether chain-of-thought actually helps, what self-consistency costs — are not this page’s business and are not adjudicated here. Runner-up spoke on the routing call was ../agentic-tooling-wiki, which owns the machinery agents are built from but holds no prompting-technique material.

Tier

T3 — the maintainer’s own repository, describing itself. The 22 notebooks are verifiable (they are there, and they were counted), but nothing about whether the techniques work, whether the code still runs against current model versions, or whether anyone finishes is evidenced here.

executable-curriculum · meta-curriculum · nir-diamant · ai-engineering-from-scratch · synthesis