context-budget-discipline

Context Budget Discipline

Context is a finite resource. Every token competes for the model's attention. The "lost-in-the-middle" phenomenon is real — content buried in 50% of the window position is least attended.

This skill applies four operations to keep the budget sane.

The Four Operations

SELECT     load context just-in-time, not all-at-once
WRITE      persist to disk; let the filesystem remember, not chat
COMPRESS   summarize older turns when window fills
ISOLATE    delegate so child work doesn't pollute parent context

SELECT — pull, don't push

Default behaviour: don't preemptively load files. Wait until they're needed.

• ❌ "Let me read all 30 .cpp files to understand the project."
• ✅ "I'll consult PROJECT-TOC.md first, then read only the 2 files the task touches."

WRITE — disk is cheap, context is expensive

Anything reusable across turns should live on disk:

• Active feature → feature_list.json
• Decisions / rationale → progress.md or docs/decisions/
• "What I just did" → session-handoff.md

Anti-pattern: re-explaining the same plan/decision in every reply.

COMPRESS — summarize old turns when window pressure rises

When context usage > 70% of model window, mid-session compaction can help. But per Anthropic Nov 2025:

"Compaction preserves continuity, but doesn't give the agent a clean slate. Context anxiety can still persist."

Flush before you compress. Before compaction/reset — or when the PostToolUse watermark fires (T1 "Context is filling" = flush now; T2 "Context is heavily filled" = wrap up via /session-end + fresh session) — write chat-only durable memory to disk first: golden-rule stubs (capturing-golden-rules), feature status, a milestone progress.md entry. After a compaction, do NOT backfill memory from the summary: reconstruct from on-disk evidence (git diff, test output, progress.md) and state uncertainty where evidence is missing.