markus41 Context Budgeting
Every token in context is a cost you pay every turn. This skill is the discipline that keeps long sessions cheap and coherent.
Token arithmetic
Rough budget on a 200K-context model:
| Consumer | Typical | Budget |
|---|---|---|
| System prompt + CLAUDE.md | 3–8 K | ≤ 10 K |
| Skills frontmatter (all active) | 2–6 K | ≤ 8 K |
| MCP tool schemas (passive) | 5–25 K | ≤ 15 K |
| Hook definitions | <1 K | <1 K |
| Conversation history | variable | ~120 K |
| Working headroom | — | ≥ 20 K |
If passive context (everything before conversation history) exceeds 30 K, optimize.
Three-tier loading (skill-creator canonical)
| Tier | What | Size | Loaded |
|---|---|---|---|
| 1 — Frontmatter | name + description |
~50 tokens | always |
| 2 — Body | SKILL.md after frontmatter |
≤ 500 lines | on activation |
| 3 — References | references/*.md, MCP KB artifacts |
unlimited | on demand |
Rule: if content doesn't change Claude's behavior every session, it belongs in tier 3.
/compact strategy
/compact condenses conversation history when it's filling up. It does NOT touch system prompt, CLAUDE.md, or skills.
When to /compact:
- Conversation > 60 K tokens.
- About to start a new phase (logical break).
- Many long tool outputs in history.
When NOT to /compact:
- Mid-debugging where specific earlier context matters.
- Right before an important decision — the lost nuance might bite.
- After a fresh
/clear(nothing to compact).
Anchor preservation: before /compact, save anything you'd lose to engram via mem_save. After compact, the post-compact-context-restoration hook re-loads memory rules and recent context.
/clear vs /compact
| Command | What it does |
|---|---|
/compact |
Summarize history, keep system prompt + CLAUDE.md |
/clear |
Wipe everything including CLAUDE.md — fresh session |
Use /clear between unrelated tasks. Use /compact mid-task.
Subagent delegation for context pressure
When a single subtask would consume too much context (e.g. scanning a huge codebase), delegate to a subagent. The agent gets a fresh window; you only see its report.
Agent({
description: "Inventory all API endpoints",
prompt: "Scan src/api/ recursively, list every route with method, auth requirement, and response schema. Report under 500 words."
})
Net: the scan consumes agent tokens; your session only pays for the 500-word summary.
Session analytics
Track cost patterns:
| Metric | Target |
|---|---|
| Passive context (pre-conversation) | ≤ 30 K |
| Cost per turn | ≤ $0.05 (Sonnet), ≤ $0.15 (Opus) |
| Tools loaded but never called per session | 0 (prune unused MCPs) |
| /compact frequency | every 20–30 exchanges on long sessions |
| Subagent delegation rate | ~1 per 5 substantive turns |
Session analytics isn't automatic — periodically review what's consuming tokens and trim.
MCP delegation
| Need | Tool |
|---|---|
| Settings schema for compact config | cc_docs_settings_schema |
| Compact/budget troubleshooting | cc_docs_troubleshoot("context") |
| Model cost for budgeting | cc_docs_model_recommend(task, budget) |
Anti-patterns
- Loading every skill body "just in case" → tier-2 explosion.
- Massive CLAUDE.md with all conventions inline → routing fails + budget bloat.
- Never compacting → session hits wall, Claude forgets the plan.
- Compacting mid-critical-reasoning → loses the thread.
- Global MCPs for project-specific work → 15 K tokens for tools you never call.
Reference
- compact-strategies.md — when to compact, anchor checklist, post-compact recovery