Summary

Use large context for orientation and mapping, then switch to a much smaller working set for precise execution.

Key points

• Wide context is good for repo mapping, transcript digestion, and architecture orientation.
• Narrow context is better for final code changes, subtle bug diagnosis, and exact reasoning.
• This pattern preserves the value of broad context without paying the full quality penalty during execution.
• Treat 1M context as an exceptional planning/orientation mode, not the default operating posture.

Operational relevance

This should become a default pattern in any serious agentic coding workflow using GPT-5.4 or similar long-context models.

Open questions

• Should the task-card schema include an explicit context tier field?
• Which current workflows cross the threshold where two-pass handling is mandatory?

Sources

• wiki/sources/gpt54-1m-context-window-operating-policy.md