The recipe maps to the literal instruction text: phrasing, examples, few-shots, role priming, formatting rules. The kitchen maps to the surrounding machinery: the retriever and its rerank pipeline, the memory store and its read/write policy, the tool registry and how their outputs thread into context, the conversation summarizer, the output schema, and the budget logic that trims when the window fills. Same recipe in a different kitchen produces a different dish.
The recipe and kitchen analogy is a useful way to make the prompt vs context engineering distinction concrete.
The **recipe** is the literal text the model reads. It includes:
- The system prompt's wording and structure.
- Role priming ('you are an expert legal analyst').
- Few-shot examples and chain-of-thought scaffolding.
- Output formatting instructions in prose.
- The phrasing of the user task.
The **kitchen** is the runtime machinery that decides what shows up on each call. It includes:
- The retriever: which chunks come back for a given query.
- The rerank pipeline: how the retrieved chunks get reordered or filtered.
- The memory store and its read/write policy: what gets persisted, what gets recalled.
- The tool registry and tool-output routing: which tools are available, how their results land in context.
- The conversation summarizer: when and how older turns get compacted.
- The output schema: how the response shape is enforced (Pydantic, JSON Schema, or prose).
- The budget logic: what gets evicted when the window fills.
The two compose. A great recipe in a broken kitchen produces a bad dish. A great kitchen with a vague recipe also produces a bad dish. Most production LLM quality work is kitchen work, even though the recipe is the visible artifact.
**The recipe is the instruction; the kitchen is everything that decides what ingredients land on the prep table.** The analogy maps the two halves of LLM application work onto two visibly different artifacts.
### The recipe, what prompt engineering owns
- **System prompt** wording, structure, numbered rules, role priming.
- **Few-shot examples** embedded in the prompt.
- **Chain-of-thought scaffolding** ('let's think step by step').
- **Output formatting** when expressed in prose.
- **Task phrasing** in the user turn.
All of this is text that ships verbatim to the model. It is human-readable, version-controllable as a single artifact, fast to iterate on.
### The kitchen, what context engineering owns
- **The pantry**: documents and embeddings in the vector store, refreshed by the indexing pipeline.
- **The shopping list**: the retrieval query and the retrieval policy (top-k, hybrid search, MMR).
- **The sous chef**: the reranker (Cohere Rerank v3, Voyage Rerank-2, BGE-Reranker for self-hosted).
- **The pantry organizer**: the memory store (Zep, Mem0, MemGPT/Letta, native session state) and its write/eviction rules.
- **The tool drawer**: tool definitions, MCP server registry, and the routing of tool outputs through summarizers before re-entry.
- **The leftover-management strategy**: the conversation summarizer that compacts older turns above a budget threshold.
- **The plating standard**: the output schema and its enforcement backend (constrained decoding, Pydantic+retry).
- **The kitchen capacity rules**: budget allocation, eviction order, output token reservation, prompt caching configuration.
All of this is composition logic, code, and infrastructure. It changes how the recipe gets executed on every call.
### Why the kitchen is most of the work in production
A team's first few months are usually spent on the recipe. The system prompt gets refined, examples get added, the role gets sharpened.
Quick reference
TL;DR
Recipe is the instruction text (wording, examples, role priming, formatting).
Think of a recipe as the page in the cookbook. Mix two cups of flour, beat three eggs, bake at 180. Anyone with the same page should make the same dish, in theory. The kitchen is everything around the page: the fridge, the pantry, the stove, the spice rack, the dishwasher, the prep table, the timer, the trash bin. The same recipe in a kitchen with a broken oven gives you a different dish than the same recipe in a fully stocked professional kitchen. With LLMs, prompt engineering is the recipe page. Context engineering is the kitchen. Two teams can use the same recipe and ship very different products because their kitchens are very different.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example.
Teaching a casual user how to interact with ChatGPT End users only see the recipe surface; the kitchen abstraction is more than they need
Single-shot generation tasks with no retrieval and no memory The kitchen is mostly trivial; almost all the work really is in the recipe for that workload
Pure prompt-engineering tutorials aimed at writers and content teams The kitchen abstraction can confuse the audience; keep the framing on the recipe for that material
Quick scan · 2–4 min
Everything important, quickly.
2 minutes: state the mapping (recipe = instruction text, kitchen = retrieval + memory + tools + summarizer + schema + budget), name at least three kitchen elements, and explain why most production quality is kitchen work.
Side-by-side
Kitchen element
LLM counterpart
Owned by
Recipe page
System prompt + examples + role priming
Prompt author / domain expert
Pantry
Vector store + indexed corpus
Data engineering
Shopping list
Retrieval query + top-k policy
Platform engineering
Sous chef
Reranker (Cohere, Voyage, BGE)
Platform engineering
Leftover system
Conversation summarizer
Platform engineering
Plating standard
Output schema (Pydantic + structured outputs)
API engineering
Real-world examples
Real products, models, and research that use this idea.
1Anthropic Claude Code's CLAUDE.md is the recipe; its tool registry, file-reading routing, compaction, and prompt-caching configuration are the kitchen.
2GitHub Copilot's system prompt is small and stable; the actual product quality comes from the kitchen: code retrieval across the repo, IDE integration, language-aware context selection.
3Cursor's background agents have a relatively short recipe; the kitchen includes file indexing, semantic code search, MCP servers, compaction, and a model ladder.
4Notion AI Q&A's recipe is short; the kitchen is per-workspace retrieval over user docs, page-level access control, and tier-aware budgets.
5Glean and Hebbia enterprise search have small recipes by comparison to their kitchens; most quality work is on retrieval pipelines, rerank, and document parsing infrastructure.
Interviewer follow-ups
What an interviewer would ask next. Try answering before peeking at the approach.
QHow does the analogy hold up for agentic systems with sub-agents?
Sign in or go Pro to see the approach for this follow-up.
QWhere does prompt caching fit in the analogy?
Sign in or go Pro to see the approach for this follow-up.
QWhat part of the analogy gets stretched too far?
Sign in or go Pro to see the approach for this follow-up.
Don't say this
Red flags and common mistakes that signal junior thinking. Click to expand.
Most common mistake
Treating prompt engineering as the whole job because that is the visible artifact. The kitchen is most of where production quality is decided, and it is invisible to the user.