How-to

Context Budget Playbook

Keep long AI sessions focused by loading the smallest useful working set and refreshing stale context deliberately.

Core conceptAnthropicOpenAI
CurrentNext review due 2026-10-21Content version 0.30.0
Format
Playbook
Level
Intermediate
Audience
Practitioner, Developer
Owner
Project42 Editorial
Review cadence
Every 90 days
Prerequisites
Experience with multi-turn AI work
01

Build a working set, not a document pile

Give the model the instructions, source material, examples, and current state needed for the next decision. Large unrelated attachments compete for attention and make it harder to identify the source of truth.

Summarize stable decisions separately from raw exploration. Keep exact source text only when wording, code, or evidence must remain precise.

02

Refresh when history becomes stale

Long sessions accumulate rejected ideas, superseded requirements, and tool output. Create a short handoff containing the goal, decisions, current state, verification, and next action before starting a clean session.

Do not compress away unresolved risks or user feedback. A useful handoff preserves why a decision was made, not only the final choice.

Clean-session handoff
text
Goal: [CURRENT OUTCOME]
Decisions: [CHOICES AND WHY]
Current state: [WHAT EXISTS NOW]
Verified: [CHECKS THAT PASSED]
Open risks: [UNCERTAINTY TO PRESERVE]
Next action: [ONE CONCRETE STEP]