The post has been translated automatically. Original language: English
Prompt engineering hasn't disappeared — it became a smaller part of something larger. Context engineering is the practice of deciding what an AI agent actually knows at the moment it acts: which definitions, which history, which prior results land in front of the model before it responds. This matters because the two disciplines have different owners. Wording a prompt is a task for whoever uses the tool. Deciding what's actually true about your business, and keeping that current where an agent can read it, is a leadership-level job that in most companies nobody is currently assigned.
How the work evolved
- Ask and answer. Output quality depended entirely on how you phrased the request.
- Saved prompts. Working formulations got written down and reused as a shared library.
- Goals instead of steps. You specify an outcome and the agent works out the process itself.
- Automatic context. The agent pulls from memory and tools at the moment of each decision, rather than waiting to be told.
- Curation. Once retrieval is automatic, the real question is what's available to retrieve and what gets selected — this is where the work sits now.
Why more context isn't better context
A widely cited study testing 18 different AI models found performance grows increasingly unreliable as input length grows, even on simple tasks. A separate research paper found models use information best at the start or end of what they're given, and worst when it's buried in the middle. An agent handed your entire data stack often performs worse than one handed the correct, curated slice of it.
What actually decides output quality
- Core definitions — what counts as a qualified lead or a successful outcome, and at what exact point.
- Who you're actually targeting right now, not who the strategy deck said last quarter.
- The real economics of the business, rather than whatever metric is easiest to optimize toward.
- The record of what's already been tried and failed — the one almost nobody wires in, and the one that matters most.
A quick way to test your own tools
- Ask the same definition question to every AI tool your team uses, without steering it. Different answers mean different systems running on different versions of your business.
- Ask for a recommendation in an area where something already failed. If the tool confidently proposes it again, that history isn't in its context.
- Take one number from an AI-generated report and try to trace it to its source. If nobody can, it's a plausible sentence, not analysis.
The fix is rarely a better prompt or a bigger model. It's putting core definitions, current priorities, real economics, and a record of past failures somewhere that's kept current and that tools actually read from.
Read the full article here.
Prompt engineering hasn't disappeared — it became a smaller part of something larger. Context engineering is the practice of deciding what an AI agent actually knows at the moment it acts: which definitions, which history, which prior results land in front of the model before it responds. This matters because the two disciplines have different owners. Wording a prompt is a task for whoever uses the tool. Deciding what's actually true about your business, and keeping that current where an agent can read it, is a leadership-level job that in most companies nobody is currently assigned.
How the work evolved
- Ask and answer. Output quality depended entirely on how you phrased the request.
- Saved prompts. Working formulations got written down and reused as a shared library.
- Goals instead of steps. You specify an outcome and the agent works out the process itself.
- Automatic context. The agent pulls from memory and tools at the moment of each decision, rather than waiting to be told.
- Curation. Once retrieval is automatic, the real question is what's available to retrieve and what gets selected — this is where the work sits now.
Why more context isn't better context
A widely cited study testing 18 different AI models found performance grows increasingly unreliable as input length grows, even on simple tasks. A separate research paper found models use information best at the start or end of what they're given, and worst when it's buried in the middle. An agent handed your entire data stack often performs worse than one handed the correct, curated slice of it.
What actually decides output quality
- Core definitions — what counts as a qualified lead or a successful outcome, and at what exact point.
- Who you're actually targeting right now, not who the strategy deck said last quarter.
- The real economics of the business, rather than whatever metric is easiest to optimize toward.
- The record of what's already been tried and failed — the one almost nobody wires in, and the one that matters most.
A quick way to test your own tools
- Ask the same definition question to every AI tool your team uses, without steering it. Different answers mean different systems running on different versions of your business.
- Ask for a recommendation in an area where something already failed. If the tool confidently proposes it again, that history isn't in its context.
- Take one number from an AI-generated report and try to trace it to its source. If nobody can, it's a plausible sentence, not analysis.
The fix is rarely a better prompt or a bigger model. It's putting core definitions, current priorities, real economics, and a record of past failures somewhere that's kept current and that tools actually read from.
Read the full article here.
Share
