What Context Engineering Is — Karpathy's Renaming

2025-06-25, Karpathy posts on X: "+1 for 'context engineering' over 'prompt engineering'. People associate prompts with short task descriptions you'd give an LLM in your day-to-day use. When in every industrial-strength LLM app, context engineering is the delicate art and science of filling the context window with just the right information for the next step."

— @karpathy, 2025-06-25

That tweet was a relay handoff from Shopify CEO Tobi Lutke's post 7 days earlier — "I really like the term 'context engineering' over prompt engineering" (@tobi, 2025-06-18). Once an LLM app stuffs 50K to 200K tokens of context per call, the prompt is just a small slice of the whole thing.

Pick Your Path First

Jump based on the engineering problem:

Where Prompt Ends and Context Begins

Tobi Lutke's definition — "the art of providing all the context for the task to be plausibly solvable by the LLM".

Prompt is the few lines you typed. Context is every token the LLM actually sees:

1. System instruction — the system prompt set in the API
2. Tool definitions — schema for every function calling / MCP tool
3. Memory / chat history — earlier turns (or summaries)
4. Retrieved context — RAG-fetched docs, vector passages, scraped pages
5. User input — what the user typed this turn

Prompt engineering writes item 5; context engineering decides how 1-5 stack into a 200K window without blowing it up.

JR's omni-report runs 17 routines — "JR AI Visibility Weekly" sweeps 20 real student queries weekly to check whether AI engines recommend JR Academy. Its prompt is doing 4 context engineering moves:

# 真实节选：omni-report/JR AI Visibility Weekly routine prompt
# tested: 2026-04-26 · model: claude-sonnet-4-6

【Phase 1：写骨架 + commit】
Write `ai-visibility/$DATE.md` 骨架（10 个 _TBD_ 占位）
commit + push: feat(ai-visibility): scaffold $DATE

【Phase 2：4 batch × 5 query × 2 layer】
每 batch 处理 5 个 query → Edit 对应表格 → commit + push
（避免 stream idle timeout）

• Context budget — task split into 6 phases with forced commits, output never piles past stream timeout
• Context selection — use ls to push upstream paths into context, don't let LLM guess
• Context scaffolding — skeleton file first (10 TBD placeholders) gives Edits structural anchors
• Context isolation — each batch commits independently, earlier mistakes don't pollute later

None are prompt wording. All structural decisions.

Why This Term Only Showed Up in 2025

ChatGPT shipped at the end of 2022. It took two and a half years for someone to say "we should rename this" — engineering reality shifted. Three things stacked up:

Thing 1: RAG went mainstream (2023-2024)

LangChain shipped v0.1 in January 2024, LlamaIndex same time. RAG turned LLM input from "a paragraph" into "a paragraph + 5 retrieved passages + summary of last 3 turns". Stanford's Lost in the Middle in July 2023 (arXiv:2307.03172, Liu et al.) proved LLM attention drops hard in the middle — docs in the prompt ≠ model uses them.

Thing 2: Agents took off (2024)

Anthropic dropped Building Effective Agents on 2024-12-20, breaking agentic systems into five patterns: prompt chaining, routing, parallelization, orchestrator-workers, evaluator-optimizer. The engineering challenge behind each is "make sure each step only sees what it should see".

Thing 3: Long-context models (2024-2025)

Claude 3.5 gave us 200K context windows. Gemini 1.5 Pro went straight to 1M. But being able to stuff isn't the same as needing to — stuffing too much drops retrieval accuracy (Lost in the Middle gets worse at 100K), cost climbs linearly, latency degrades. More tokens means token allocation became the new engineering problem.

Three threads converged by mid-2025, and Tobi and Karpathy said nearly the same line within 7 days in June. Simon Willison's 2025-06-27 post recaps — not a naming gimmick, 2 years of engineer muscle memory finally getting a name.

Prompt Engineering vs Context Engineering — Not Replacement, It's a Layer Above

A lot of people misread this as "prompt eng is obsolete". Prompt is part of context. Their relationship:

Prompt engineering solves the "communication problem". Context engineering solves the "system problem". First is like writing an email. Second is like designing the email protocol.

Why "Engineering" Instead of "Design"

Other people call this "context design" or "prompt craftsmanship". Karpathy deliberately picked the word engineering — that's a signal:

• Measurable — context quality runs through an eval set of 1000 queries, not "I think it looks good"
• Reusable — context selection strategy is code, drops into the next project
• Can break — one deploy tweaks retrieval threshold, accuracy drops 15% — incident, not taste
• Trade-offs explicit — 5 more passages vs cost 2x vs latency +800ms, pick one

Anthropic's prompt caching docs — the 5-minute TTL is engineering trade-off. A cache hit saves 90% on cost, but past 5 minutes you rebuild it.

What's Coming Across the Next 9 Chapters

Every chapter unpacks a real engineering problem with code + JR omni-report cases.

Takeaway

Prompt is "what you say". Context is "what the LLM sees". Between "what you say" and "what the LLM sees" sits 4 more layers — system prompt, tool definitions, memory, retrieval. Those 4 layers are where 80% of the engineering work in a production LLM app lives.

---

References

1. Karpathy. (2025-06-25). Tweet on context engineering.
2. Tobi Lutke. (2025-06-18). Tweet on context engineering.
3. Anthropic. (2024-12-20). Building Effective Agents.
4. Liu et al. (2023-07-06). Lost in the Middle. arXiv:2307.03172.
5. Simon Willison. (2025-06-27). Context engineering.
6. Anthropic. Prompt caching docs.

Production case: JR Academy omni-report — context engineering across 17 routines.