All articlesPrompt Engineering
Context Engineering Beats Longer Prompts
Bigger context windows made prompt bloat cheap and quality worse. What matters is what you put in the window, and in what order.
Sri Raman30 July 20267 min read
The shift in the last two years is not that prompts got better — it is that context got bigger, and teams filled it indiscriminately. Context engineering is the discipline of deciding what earns a place in the window.
Budget the window like a page layout
- Instructions: stable, short, and never duplicated.
- Task input: the actual thing to work on, clearly delimited.
- Evidence: only retrieved chunks that could change the answer.
- Format contract: the exact output shape, stated last.
Position matters more than people admit
Models weight the beginning and the end of a long context more reliably than the middle. Put the non-negotiables — role, constraints, output format — at the edges, and bury nothing important in the centre of a 30k-token dump.
Delimit everything
<instructions>
...stable rules...
</instructions>
<evidence>
[1] source: policy.pdf p.4
...
</evidence>
<task>
{{user_request}}
</task>Tagged blocks make it possible to reference sources by number, strip a section programmatically, and diff two prompts without guessing where one part ends.
Prune on a schedule
- 1Log the full context for a sample of real requests.
- 2Remove one block and re-run your evaluation set.
- 3If quality holds, that block was decoration — delete it permanently.
A prompt you cannot explain line by line is a prompt you cannot debug.
