- Most teams try to fix AI output by rewriting the prompt. In production systems, the bigger lever is context: what the model can see when it decides, and how that information is selected, ordered, and kept fresh.
- We learned this running AI across three jobs at Elevarus: writing content, reporting on performance, and operating a fleet of marketing agents. Our worst failures were context failures, not prompt failures.
- More context is not better context. Good context is curated, not accumulated: include what is relevant, cut the noise, order what remains, and expire what is stale.
- The fix is an order of operations. Engineer the context first. Tune the prompt second.
Most teams still think the fix for bad AI output is a better prompt. Pick sharper words, add a clever framework, find the magic phrasing, and the model finally behaves.
Sometimes that is true. In a quick one-off chat, wording moves the needle.
But in the systems we actually run at Elevarus, prompting is rarely the lever that matters most. These are the systems that generate leads, write content, and report on performance every day. The bigger lever, by a wide margin, is context. By context we mean what the model can see when it makes a decision, and how that information is selected, structured, and kept fresh.
We learned this the slow way. We spent real time rewriting instructions to fix problems that wording was never going to fix. Once we stopped obsessing over the prompt and started engineering the context, the quality of everything downstream changed.
This is not just our private vocabulary. Anthropic, whose Claude models we build on, calls this discipline context engineering and frames it as the natural progression beyond prompt engineering once models start using tools and working across real tasks. Here is what that shift looks like in a real performance-marketing operation.
Quick answers:
- What is context engineering?
- What is the difference between context and prompt engineering?
- Is prompt engineering dead?
- Why does more context not always improve AI output?
- How do you start engineering context?
- How does this apply to marketing and lead generation?
Prompting gets the credit because you can watch it work
Prompt tuning is seductive for one reason. The feedback loop is instant. You change a word, you watch the answer change, and it feels like you are in control of quality.
That illusion breaks the moment AI leaves the chat box and enters a workflow. A lead-routing decision. A campaign report. A published article. We watched strong prompts produce weak output for reasons that had nothing to do with phrasing:
- The model pulled the wrong source document.
- The brief arrived missing the campaign context that actually mattered.
- The window was flooded with irrelevant data, so the important facts got buried.
- Stale numbers from an earlier step quietly skewed the answer.
- A tool dumped raw JSON the model had to fight through.
In every one of those cases, rewriting the instruction was paint on rust. The fix was always upstream, in what the model was reading, not in how we asked.
What context actually means in our work
People use the word loosely. In practice, context is everything that shapes the model’s next decision. In a real system, that is a lot more than the user’s message.
When one of our agents drafts a lead-gen article, its context includes the target keyword’s live search landscape. It includes the competitor pages actually ranking. It includes our brand voice rules and the compliance limits for that vertical, what we can and cannot claim about insurance or financial leads. It includes the internal pages it should link to and the structure the piece has to follow. The instruction, “write a buyer’s guide,” is the smallest part of that.
So when we say our AI performs well, we do not mean it reads long text. We mean something narrower. It does good work when the right information is present, arranged clearly, kept relevant, and refreshed at the right moment. That is a design problem, not a wording problem.
A useful mental model comes from Andrej Karpathy, who helped popularize the term. He calls context engineering the delicate art and science of filling the context window with just the right information. A simple way to picture it: the model is the processor, and its context window is the working memory you have to load with exactly the right data for the task. Frame it that way and the work stops being about clever phrasing and starts being about what you load in, and in what order.
Three places engineering context changed how we operate
Content creation
Our content engine does not win because we found a perfect “write an SEO article” prompt. It wins because of the work that happens before the model writes a single sentence.
By then the system has already done the hard part. It researched the live web for that exact buyer query. It pulled the real questions buyers ask and the benchmarks that matter. It filtered out the noise. Then it handed the model a curated evidence base plus our voice and compliance rules, in that order.
Swap a polished prompt onto a thin, generic context and you get a confident, well-written article that is vague or wrong. Give the model a strong base of real information and a simpler instruction, and the output is far better. We see it every week. The post that reads like it gets the vertical is not the product of clever wording. It is the product of better context. The same discipline sits underneath all of our lead generation content.
Reporting
Lead-gen reporting is where bad context does the quietest damage. It is tempting to hand the model the entire export, every campaign, every call record, every raw field, and ask for insights. What comes back is mush. Weak prioritization. Noise mistaken for signal. Confident conclusions built on the wrong column.
What works is the opposite. We feed the model a curated, labeled slice. Cost per lead and cost per call, computed cleanly. The comparison framed. The irrelevant fields stripped. The exact numbers it should reason over highlighted.
Same model, same goal. But the framing changes everything. “Here is a data dump, find something” gets you a report you have to double-check. “Here is the decision-relevant evidence, structured” gets you one you trust. This matters most on paid channels, where a misread number can move real budget. It is the same logic that drives how we report on Google Ads management.
Running the whole operation
Behind the scenes, Elevarus runs marketing as a set of specialized AI agents. One writes. One edits. One handles technical SEO. Others do site work and strategy. The thing that makes that fleet reliable is not the cleverness of any single prompt.
It is context discipline. Each agent works from a curated memory of what it has already done, what the others are doing, and what the current goal is. That memory is deliberately bounded so it does not drown the agent in its own history.
Our biggest failures were not bad instructions. They were context failures. An agent operating on stale state. A session that accumulated so much history it stalled. The fixes were context fixes every time. Refresh the state. Bound the memory. Hand off clean.
More context is not better context
Here is the trap that catches most teams. If some context helps, surely more context helps more.
It does not. We have degraded our own output by over-stuffing the window. Every document, every prior turn, every tool result, all at once. The result is slower reasoning, worse prioritization, and more chances for the model to trip over a contradiction we handed it.
This is not just a feel. Anthropic describes the underlying effect plainly: as the number of tokens in the context window grows, the model’s ability to accurately recall information from that context goes down. They call the goal finding the smallest possible set of high-signal tokens that get the outcome you want.
Good context is not large. It is useful. That means deciding, on purpose:
| Decision | What it means |
|---|---|
| Include | Add what is decision-relevant |
| Exclude | Leave out what distracts |
| Order | Arrange what remains so the model weights it correctly |
| Expire | Drop what has gone stale instead of letting it pile up |
Curation, not accumulation. That single shift fixed more quality problems for us than any prompt rewrite ever did.
Where most teams go wrong
If your AI output is inconsistent, the instinct is to blame the prompt. Usually it is one of these instead:
- Retrieval is weak, so the model is quietly guessing. A guess dressed in good prose looks like an answer.
- Memory is stale, so a decision is made on yesterday’s numbers.
- Tool output is raw, so the model is parsing instead of reasoning.
- The window is overloaded, so the one fact that mattered competes with fifty that did not.
- The task should have been split, but it was crammed into a single step.
A better prompt cannot fix any of those. At best it makes the wrong answer sound more confident, which is worse than no progress, because now it is harder to catch. As one widely shared engineering write-up put it, prompt engineering walked so context engineering could run. The systems that hold up in production are the ones that treat the information environment as the real work.
A simple order of operations: context first, prompt second
When output quality matters, we work the context before we touch the wording. The checklist we actually use:
First, the context:
- Do we have the right source material?
- Is what we retrieved actually relevant and current?
- Are we passing too much noise?
- Is the information ordered so the model weights the right part?
- Are tool outputs cleaned and labeled, or raw?
- Is stale state still hanging around?
- Would this work better as two or three steps instead of one?
Then, the prompt:
- Is the instruction clear?
- Is the task narrow enough?
- Are the constraints and output format explicit?
Context first. Prompt second. That order is the whole lesson.
The bottom line
Our AI does not outperform a better prompt because we found smarter words. It outperforms because, by the time the model decides anything, we have already built the right informational world around it. The right facts, in the right structure, at the right moment.
Prompting still matters. But for any team running AI in production, generating leads, reporting on spend, publishing content at scale, context is the bigger lever now. If you want better results, stop asking how to write a stronger prompt. Start asking how to engineer a stronger environment around the model.
That is where the real performance gains live. It is where ours came from. If you want help putting AI to work on your own pipeline, talk to our team.
Frequently Asked Questions
What is context engineering?
Context engineering is the practice of curating and maintaining the information a model sees when it makes a decision. That includes the system instructions, retrieved documents, tool outputs, prior messages, and memory, not just the prompt. The goal is to give the model the right information, in the right structure, at the right moment, rather than the most information.
What is the difference between context and prompt engineering?
Prompt engineering is about wording: finding the instruction that gets a good answer. Context engineering is about the whole information environment around that instruction. In a quick chat, wording matters most. In a production system that uses tools, memory, and live data, the context is usually the bigger lever, because a perfect prompt still fails on the wrong inputs.
Is prompt engineering dead?
No. Prompt engineering still matters, especially for one-off tasks and for making instructions clear and constrained. What has changed is the priority. Once AI runs inside real workflows, the wording is a smaller part of the job than the information you feed in. Treat the prompt as the last step, not the first.
Why does more context not always improve AI output?
Because models lose accuracy as the context window fills up. There is a name for the effect, context rot. As Anthropic describes it, the more tokens you add, the harder it is for the model to recall the ones that matter. Stuffing in every document and every past message buries the signal in noise, slows reasoning, and introduces contradictions. Useful context beats large context.
How do you start engineering context?
Work an order of operations. Before touching the prompt, check the inputs. Do you have the right source material? Is it current and relevant? Are you passing noise, or is the important part ordered first? Are tool outputs cleaned up, and is stale state gone? Only then tune the instruction. Context first, prompt second.
How does this apply to marketing and lead generation?
In marketing, context discipline shows up in two places. In content, the system researches the live query and hands the model a curated evidence base before it writes. In reporting, you feed the model a clean, labeled slice of performance data rather than a raw export. Both produce work you can trust, because the model is reasoning over the right facts instead of guessing.
