The next is Half 1 of three from Addy Osmani’s unique submit “Context Engineering: Bringing Engineering Self-discipline to Components.”
Context Engineering Ideas:
To get the perfect outcomes from an AI, it’s worthwhile to present clear and particular context. The standard of the AI’s output instantly is determined by the standard of your enter.
The right way to enhance your AI prompts:
- Be exact: Imprecise requests result in obscure solutions. The extra particular you might be, the higher your outcomes shall be.
- Present related code: Share the particular information, folders, or code snippets which might be central to your request.
- Embrace design paperwork: Paste or connect sections from related design docs to offer the AI the larger image.
- Share full error logs: For debugging, at all times present the whole error message and any related logs or stack traces.
- Present database schemas: When working with databases, a screenshot of the schema helps the AI generate correct code for information interplay.
- Use PR suggestions: Feedback from a pull request make for context-rich prompts.
- Give examples: Present an instance of what you need the ultimate output to appear like.
- State your constraints: Clearly listing any necessities, resembling libraries to make use of, patterns to comply with, or issues to keep away from.
Immediate engineering was about cleverly phrasing a query; context engineering is about setting up a whole data surroundings so the AI can resolve the issue reliably.
“Immediate engineering” turned a buzzword primarily that means the ability of phrasing inputs to get higher outputs. It taught us to “program in prose” with intelligent one-liners. However outdoors the AI group, many took immediate engineering to imply simply typing fancy requests right into a chatbot. The time period by no means absolutely conveyed the true sophistication concerned in utilizing LLMs successfully.
As functions grew extra advanced, the constraints of focusing solely on a single immediate turned apparent. One evaluation quipped: Immediate engineering walked so context engineering may run. In different phrases, a witty one-off immediate may need wowed us in demos, however constructing dependable, industrial-strength LLM techniques demanded one thing extra complete.
This realization is why our discipline is coalescing round “context engineering” as a greater descriptor for the craft of getting nice outcomes from AI. Context engineering means setting up the complete context window an LLM sees—not only a brief instruction, however all of the related background information, examples, and steerage wanted for the duty.
The phrase was popularized by builders like Shopify’s CEO Tobi Lütke and AI chief Andrej Karpathy in mid-2025.
“I actually just like the time period ‘context engineering’ over immediate engineering,” wrote Tobi. “It describes the core ability higher: the artwork of offering all of the context for the duty to be plausibly solvable by the LLM.” Karpathy emphatically agreed, noting that “individuals affiliate prompts with brief directions, whereas in each severe LLM software, context engineering is the fragile artwork and science of filling the context window with simply the correct data for every step.”
In different phrases, real-world LLM apps don’t succeed by luck or one-shot prompts—they succeed by rigorously assembling context across the mannequin’s queries.
The change in terminology displays an evolution in strategy. If immediate engineering was about arising with a magical sentence, context engineering is about writing the complete screenplay for the AI. It’s a structural shift: Immediate engineering ends when you craft an excellent immediate, whereas context engineering begins with designing complete techniques that herald reminiscence, information, instruments, and information in an organized approach.
As Karpathy defined, doing this effectively includes all the pieces from clear process directions and explanations, to offering few-shot examples, retrieved information (RAG), presumably multimodal information, related instruments, state historical past, and cautious compacting of all that right into a restricted window. Too little context (or the incorrect type) and the mannequin will lack the data to carry out optimally; an excessive amount of irrelevant context and also you waste tokens and even degrade efficiency. The candy spot is non-trivial to seek out. No marvel Karpathy calls it each a science and an artwork.
The time period context engineering is catching on as a result of it intuitively captures what we truly do when constructing LLM options. “Immediate” seems like a single brief question; “context” implies a richer data state we put together for the AI.
Semantics apart, why does this shift matter? As a result of it marks a maturing of our mindset for AI improvement. We’ve discovered that generative AI in manufacturing is much less like casting a single magic spell and extra like engineering a whole surroundings for the AI. A one-off immediate would possibly get a cool demo, however for sturdy options it’s worthwhile to management what the mannequin “is aware of” and “sees” at every step. It usually means retrieving related paperwork, summarizing historical past, injecting structured information, or offering instruments—no matter it takes so the mannequin isn’t guessing at nighttime. The result’s we now not consider prompts as one-off directions we hope the AI can interpret. We expect by way of context pipelines: all of the items of data and interplay that set the AI up for achievement.
For example, take into account the distinction in perspective. Immediate engineering was usually an train in intelligent wording (“Perhaps if I phrase it this fashion, the LLM will do what I would like”). Context engineering, against this, feels extra like conventional engineering: What inputs (information, examples, state) does this technique want? How do I get these and feed them in? In what format? At what time? We’ve primarily gone from squeezing efficiency out of a single immediate to designing LLM-powered techniques.
What Precisely Is Context Engineering?
Context engineering means dynamically giving an AI all the pieces it must succeed—the directions, information, examples, instruments, and historical past—all packaged into the mannequin’s enter context at runtime.
A helpful psychological mannequin (instructed by Andrej Karpathy and others) is to consider an LLM like a CPU, and its context window (the textual content enter it sees without delay) because the RAM or working reminiscence. As an engineer, your job is akin to an working system: load that working reminiscence with simply the correct code and information for the duty. In follow, this context can come from many sources: the consumer’s question, system directions, retrieved information from databases or documentation, outputs from different instruments, and summaries of prior interactions. Context engineering is about orchestrating all these items into the immediate that the mannequin finally sees. It’s not a static immediate however a dynamic meeting of data at runtime.
Let’s break down what this includes:
- It’s a system, not a one-off immediate. In a well-engineered setup, the ultimate immediate the LLM sees would possibly embrace a number of elements: e.g., a job instruction written by the developer, plus the newest consumer question, plus related information fetched on the fly, plus maybe a couple of examples of desired output format. All of that’s woven collectively programmatically. For instance, think about a coding assistant AI that will get the question “How do I repair this authentication bug?” The system behind it’d mechanically search your codebase for associated code, retrieve the related file snippets, after which assemble a immediate like: “You might be an professional coding assistant. The consumer is dealing with an authentication bug. Listed below are related code snippets: [code]. The consumer’s error message: [log]. Present a repair.” Discover how that ultimate immediate is constructed from a number of items. Context engineering is the logic that decides which items to tug in and the best way to be a part of them. It’s akin to writing a perform that prepares arguments for an additional perform name—besides right here, the “arguments” are bits of context and the perform is the LLM invocation.
- It’s dynamic and situation-specific. In contrast to a single hard-coded immediate, context meeting occurs per request. The system would possibly embrace completely different information relying on the question or the dialog state. If it’s a multi-turn dialog, you would possibly embrace a abstract of the dialog thus far, relatively than the complete transcript, to save lots of area (and sanity). If the consumer’s query references some doc (“What does the design spec say about X?”), the system would possibly fetch that spec from a wiki and embrace the related excerpt. Briefly, context engineering logic responds to the present state—very like how a program’s habits is determined by enter. This dynamic nature is essential. You wouldn’t feed a translation mannequin the very same immediate for each sentence you translate; you’d feed it the brand new sentence every time. Equally, in an AI agent, you’re continually updating what context you give because the state evolves.
- It blends a number of kinds of content material. LangChain describes context engineering as an umbrella that covers no less than three sides of context: (1) Tutorial context—the prompts or steerage we offer (together with system function directions and few-shot examples), (2) Data context—area data or information we provide, usually by way of retrieval from exterior sources, and (3) Instruments context—data coming from the mannequin’s surroundings by way of instruments or API calls (e.g., outcomes from an internet search, database question, or code execution). A strong LLM software usually wants all three: clear directions in regards to the process, related information plugged in, and presumably the flexibility for the mannequin to make use of instruments after which incorporate the software outcomes again into its pondering. Context engineering is the self-discipline of managing all these streams of data and merging them coherently.
- Format and readability matter. It’s not simply what you embrace within the context, however how you current it. Speaking with an AI mannequin has stunning parallels to speaking with a human: In the event you dump an enormous blob of unstructured textual content, the mannequin would possibly get confused or miss the purpose, whereas a well-organized enter will information it. A part of context engineering is determining the best way to compress and construction data so the mannequin grasps what’s necessary. This might imply summarizing lengthy texts, utilizing bullet factors or headings to focus on key information, and even formatting information as JSON or pseudo-code if that helps the mannequin parse it. For example, for those who retrieved a doc snippet, you would possibly preface it with one thing like “Related documentation:” and put it in quotes, so the mannequin is aware of it’s reference materials. In case you have an error log, you would possibly present solely the final 5 traces relatively than 100 traces of stack hint. Efficient context engineering usually includes inventive data design—making the enter as digestible as doable for the LLM.
Above all, context engineering is about setting the AI up for achievement.
Keep in mind, an LLM is highly effective however not psychic—it might probably solely base its solutions on what’s in its enter plus what it discovered throughout coaching. If it fails or hallucinates, usually the foundation trigger is that we didn’t give it the correct context, or we gave it poorly structured context. When an LLM “agent” misbehaves, often “the suitable context, directions and instruments haven’t been communicated to the mannequin.” Rubbish in, rubbish out. Conversely, for those who do provide all of the related information and clear steerage, the mannequin’s efficiency improves dramatically.
Feeding high-quality context: Sensible suggestions
Now, concretely, how will we guarantee we’re giving the AI all the pieces it wants? Listed below are some pragmatic suggestions that I’ve discovered helpful when constructing AI coding assistants and different LLM apps:
- Embrace related supply code and information. In the event you’re asking an AI to work on code, present the related code information or snippets. Don’t assume the mannequin will recall a perform from reminiscence—present it the precise code. Equally, for Q&A duties embrace the pertinent information or paperwork (by way of retrieval). Low context ensures low-quality output. The mannequin can’t reply what it hasn’t been given.
- Be exact in directions. Clearly state what you need. In the event you want the reply in a sure format (JSON, particular fashion, and many others.), point out that. If the AI is writing code, specify constraints like which libraries or patterns to make use of (or keep away from). Ambiguity in your request can result in meandering solutions.
- Present examples of the specified output. Few-shot examples are highly effective. In order for you a perform documented in a sure fashion, present one or two examples of correctly documented features within the immediate. Modeling the output helps the LLM perceive precisely what you’re searching for.
- Leverage exterior information. If the duty wants area information past the mannequin’s coaching (e.g., company-specific particulars, API specs), retrieve that information and put it within the context. For example, connect the related part of a design doc or a snippet of the API documentation. LLMs are much more correct once they can cite information from offered textual content relatively than recalling from reminiscence.
- Embrace error messages and logs when debugging. If asking the AI to repair a bug, present it the complete error hint or log snippet. These usually comprise the crucial clue wanted. Equally, embrace any check outputs if asking why a check failed.
- Preserve dialog historical past (well). In a chat situation, feed again necessary bits of the dialog thus far. Typically you don’t want the complete historical past—a concise abstract of key factors or selections can suffice and saves token area. This offers the mannequin context of what’s already been mentioned.
- Don’t draw back from metadata and construction. Typically telling the mannequin why you’re giving a bit of context may also help. For instance: “Right here is the consumer’s question.” or “Listed below are related database schemas:” as prefacing labels. Easy part headers like “Person Enter: … / Assistant Response: …” assist the mannequin parse multi-part prompts. Use formatting (markdown, bullet lists, numbered steps) to make the immediate logically clear.
Keep in mind the golden rule: LLMs are highly effective however they aren’t mind-readers. The standard of output is instantly proportional to the standard and relevance of the context you present. Too little context (or lacking items) and the AI will fill gaps with guesses (usually incorrect). Irrelevant or noisy context might be simply as dangerous, main the mannequin down the incorrect path. So our job as context engineers is to feed the mannequin precisely what it wants and nothing it doesn’t.
AI instruments are rapidly shifting past chat UX to classy agent interactions. Our upcoming AI Codecon occasion, Coding for the Future Agentic World, will spotlight how builders are already utilizing brokers to construct progressive and efficient AI-powered experiences. We hope you’ll be a part of us on September 9 to discover the instruments, workflows, and architectures defining the following period of programming. It’s free to attend.
