The Annual Context Window Census finds that models can hold more than ever and humans remain capable of pasting all of it. Larger windows are useful. They are not permission to stop deciding what matters.
Capacity is not attention.
A model may accept a repository, a meeting archive, and the complete
emotional history of a variable named data. Acceptance
does not prove that every token helps. Irrelevant material can bury
constraints, introduce stale facts, and make a confident answer harder
to audit because the evidence is spread across a small civilization.
The context window remembers. Relevance still needs an editor.
| Observed condition | Department estimate | Recommended response |
|---|---|---|
| Files declared “probably relevant” | All of them | Start with the smallest sufficient set. |
| Original questions still visible after the preamble | Concerningly few | State the task first and repeat it after long evidence. |
| API keys improved by appearing in context | 0% | Remove credentials before anything else. |
Practice progressive disclosure.
Begin with the goal, relevant constraints, and the minimum evidence needed for a useful first pass. Add more context when the response exposes a real gap. This keeps the conversation inspectable and makes incorrect assumptions easier to locate.
- Write the desired outcome in one sentence.
- Include current, authoritative sources before secondary summaries.
- Remove secrets, personal data, generated files, and unrelated history.
- Label uncertainty and contradictory evidence instead of flattening it.
- After a long exchange, restate the current decision and verify it outside the chat.
The useful limit is intentionality.
Context windows will continue to grow. The scarce resource is not token capacity; it is the human attention required to select evidence, notice drift, and decide whether the answer actually solved the problem.