Claude Code Context Window Optimization

Guide by @aakashgupta: Context is the scarcest resource in Claude Code. Before the first user message is sent, 10%–16% of the context window is already consumed by system prompts, MCP servers, and custom agents.

Overhead breakdown

Cost hierarchy (lowest → highest overhead)

1. CLI — zero overhead (e.g., GitHub CLI, Vercel CLI, Firecrawl CLI)
2. API — moderate overhead
3. MCP — highest overhead

Andrej Karpathy independently confirmed this CLI > API > MCP ordering.

Practical tactics

Monitor usage

• Run /status line to see real-time context consumption.
• Set color-coded thresholds:
  • Green: < 50%
  • Orange: 50–80%
  • Red: > 80%

Prefer CLI over MCP

• Before adding an MCP server, ask whether a CLI tool can achieve the same result.
• Periodically audit existing MCP integrations for CLI replacements.

Escape double-tap

• If Claude starts drifting off-topic, press Escape twice.
• This rolls back to before the erroneous prompt and physically removes the subsequent content from context.

Pawel Huryn's Four Root Causes (2026-04-26)

Pawel Huryn documented the four root causes of excessive Claude Code token consumption after Anthropic fixed three platform bugs (v2.1.116+):

1. Cache misses

The prompt cache is the single biggest lever. Cache read costs 0.1x input price; cache write costs 1.25x (5-min TTL) or 2x (1-hour TTL). Every hit on a cached prefix resets its TTL at no extra cost.

Key rules:

• Lock tools at session start. Adding or removing a tool mid-session invalidates the cached prefix.
• Lock the model at session start. Switching models mid-session blows the cache.
• ~90% hit rate is healthy on the 5-minute default. On 1-hour TTL (API-only), it climbs to ~97-99%.

2. Context bloat

For Opus 4.7, 1M context is the default but expensive. Long sessions sprawl and auto-compact fires later than optimal.

Recommended settings:

{
  "env": {
    "CLAUDE_CODE_DISABLE_1M_CONTEXT": "1",
    "CLAUDE_AUTOCOMPACT_PCT_OVERRIDE": "80"
  }
}

This disables 1M context (fallback to 200K) and pins auto-compact at 80%.

Session moves:

• /compact at 50% or after every task. Do not wait for auto.
• /clear between unrelated work. New session = fresh prefix.
• /rewind when a turn went sideways. Cheaper than re-prompting around bad context.
• Subagents for anything that does not need the parent's reasoning.

3. Wrong model or effort

Three separate dials that burn tokens fast if misconfigured:

Set effort per prompt, not per task or session.

Model routing in CLAUDE.md:

• Sonnet session: cheaper, good when work is in-scope for Sonnet
• Opus session + delegate: pay for Opus on parent (planning, tradeoffs), delegate rest

4. Wrong input format

Some inputs are token-expensive by default:

Additional tools:

• rtk-ai/rtk: CLI proxy that strips redundant whitespace and compresses tool output (~20-30% token reduction)
• juliusbrussee/caveman: terse-output skill that drops conversational filler without affecting reasoning

Ronin 的五大上下文陷阱与四层模型路由（2026-05-13）

来源：Ronin — 如何把 AI 编程账单砍掉 80%

Ronin 将月开销从 4,200 美元降到 312 美元，核心方法是识别并消除无效上下文。

五大陷阱

四层模型路由体系

关键原则：不要用一个模型处理所有事。全用 Sonnet 折中方案可能让成本比实际需求高 6 倍。

Counterpoints & Gaps

• The exact percentages (2%, 8%, 4%) are approximate and may vary by model version and prompt length.
• Some workflows genuinely require MCP-level integration; CLI substitution is not always feature-complete.
• There is no documented guarantee that Escape double-tap permanently erases tokens from the billing context.
• Ronin 的 80% 降幅基于特定项目结构，不具有普遍代表性。但五大陷阱的定性判断与 Pawel Huryn 的框架一致。