Journal / AI
Better context beats a bigger prompt
Reliable AI workflows need a deliberate information environment, not an ever-growing conversation.
The working set matters
Anthropic’s context-engineering guidance describes context as a limited resource that needs to be curated throughout an agent’s work. Its recommendations include selective retrieval, compact summaries, and retaining the information needed for the next decision. This is useful even when a model accepts a very large input: availability does not make every document equally relevant.
Separate knowledge from instructions
For a product team, separate three things. Instructions define the task and boundaries. Reference material supplies facts and examples. Working state records what has been tried and what remains unresolved. Mixing them into one long prompt makes stale assumptions harder to detect. A retrieved document should arrive with enough source information for its claims to be checked, especially when the document can change.
Our take
A useful design exercise is to inspect one failed answer and ask what information was missing, distracting, or treated as authoritative without justification. Then change the retrieval or presentation of that information and try the same task again. Measure whether the output improves. Adding another page of instructions can conceal a data problem; designing the information flow makes the problem easier to see.
Put it into practice.
Keep a short task brief, a small set of relevant sources, and an explicit record of unresolved questions. Revisit each when the task changes.
Sources & further reading
Editorial analysis by BEMATE. Sources reviewed 6 September 2026.