Context is a Per-Feature Budget is the operational principle that treats an LLM’s context window not as a persistent workspace, but as a finite, depleting resource allocated to a single unit of work (e.g., a feature, a bug fix, or a distinct task).
The Core Concept
Long-running chat sessions suffer from Context Decay. As a conversation grows, the model’s adherence to initial system instructions, quality bars, and specific constraints degrades. The model becomes “forgetful” or “lazy,” often silently dropping requirements (measured at ~40% drop rate in some experiments).
To combat this, you must reset the context (start a new chat) for every new unit of work.
The Rule: One Unit of Work = One Context
- Scope the Work: Define a single feature or task.
- Spend the Budget: Use the context window solely to execute that task.
- Reset: Once the task is complete (or if the context heavily degrades), end the session.
- Repeat: Start a fresh context for the next unit of work.
“A new unit of work gets a fresh context.” — AI Techniques Distilled
The Prerequisite: External Memory
You cannot safely reset context if the chat history is your only memory. This principle requires shifting state from chat to durable artifacts (files).
- Before Resetting: Ensure all progress, decisions, and next steps are captured in files (e.g.,
PRD.md,todo.md,user_requests/feature_A.md). - Rehydrating: When starting the new context, read in only the relevant files needed for the new task.
Relation to Context Hygiene
This principle is a specific, actionable rule within the broader practice of Context Hygiene.
- Context Hygiene is the discipline (don’t pollute, keep it clean).
- Per-Feature Budget is the tactic (reset on every feature boundary).
References
- AI Techniques Distilled From Thousands of Hours of Real Work: Identifies “Silent Requirement Dropping” as the failure mode and “Clear Context” as the fix.
- Fundamental Knowledge SWEs in 2026 Must Have: Geoffrey Huntley emphasizes “One task per context window” and “Hit new chat all the time.”