← Writing

28 February 20268 min read

Context engineering is what matters now, not prompt engineering

The era of clever prompting is over. The skill that separates engineers who get consistent results from AI is managing what the model knows — not how you ask.

Prompt engineeringClaudeGeminiAI engineeringWorkflow

The skill that aged out

Two years ago, prompt engineering was a skill worth developing. The techniques were specific: chain-of-thought prompting, few-shot examples, role assignment, output format specification. There were courses, certifications, entire job titles.

Most of that is now obsolete.

Not because prompting does not matter — a badly phrased instruction still produces a worse output than a well-phrased one. But the marginal return on prompt cleverness has collapsed as models have gotten better. Claude, Gemini, and GPT-4o are robust to poor phrasing in a way that earlier models were not.

The skill that now drives the quality of AI-assisted work is context engineering: deliberate management of what information the model has access to when it generates a response.

What context engineering means

Context engineering is the practice of curating, structuring, and delivering the information a model needs to produce the output you want — before you ask for anything.

It is less about how you ask and more about what the model knows when you ask.

In a Claude Code session, context engineering means starting every session with a structured brief: the architecture of the system, the relevant constraints, the outcome you need, the things to explicitly avoid. Not as a prompt trick — as genuine information transfer. The model cannot do good work in an informational vacuum regardless of how cleverly you phrase the request.

In a Gemini document analysis task, it means thinking carefully about which documents to include, in what order, and whether to include a summary frame that tells the model what relationships to look for across them.

In a Manus research workflow, it means providing enough specificity about what you need — the domain, the level of detail, the form of the output — that the agent's browsing is directed rather than exploratory.

The common thread: the engineer's job before the model runs is as important as the output from the model after it runs.

The failure mode it fixes

The most common failure mode in AI-assisted engineering is the mismatch between what the model knows and what the engineer assumes it knows.

An engineer asks Claude to refactor a function. Claude produces clean code that misses a constraint the engineer considers obvious — a thread-safety requirement, a naming convention, a dependency that cannot change. The engineer reverts the change and concludes the model is unreliable.

The model was not unreliable. It was uninformed. It produced a correct output for the task as it understood it.

Context engineering is the practice of closing this gap: auditing your own assumptions about what the model knows and making them explicit before the task runs.

This reframe changes the debugging process when AI outputs are wrong. Instead of "the model got this wrong," the question becomes "what information would have made this right?" That question is almost always answerable, and the answer is almost always actionable.

What to include in context

For code generation tasks with Claude:

For long-document tasks with Gemini:

For research tasks with Manus:

The pattern across all three: context engineering is about building a mental model handoff. You are transferring enough of your understanding to the model that it can make the same decisions you would make.

The new skill hierarchy

The engineers getting the best results from AI tooling in 2026 are not the ones who learned the most prompt tricks. They are the ones who are most rigorous about information.

They think carefully before every AI-assisted task: what does the model need to know? What am I assuming it already knows? What constraints have I not stated explicitly?

They treat the context setup as the professional skill and the prompt itself as secondary. They have internalized that a well-informed model with a simple prompt outperforms an uninformed model with a clever one.

This is a mental shift that is available to any engineer willing to make it. It does not require a course or a certification. It requires the discipline to pause before sending and ask: does this model have everything it needs?

Almost always, the honest answer is no. And the gap is almost always fillable.

The diminishing returns signal

You know you have moved from prompt engineering to context engineering when your responses to bad model outputs change.

Old response to a bad output: try a different prompt, add more emphasis, rephrase the instruction.

New response: identify the missing or incorrect context that produced this output, add it, rerun.

The second response fixes the root cause. The first response is guessing.

When your instinct is to diagnose before you rephrase, you are doing context engineering. That is the switch worth making.

← All writingWork with me →