Guides · Guide
How to Save Tokens in Claude Code: Effort Levels, Context Hygiene and Skills
Token bills creep up from a few habits: max effort on easy tasks, a context window full of old output, and instructions you load every session. Here is what effort levels actually change and the habits that cut usage without making the agent dumber.

Most people who blow through their Claude usage are not doing anything exotic. They leave effort on the highest setting for renaming a variable, let one session run for six hours, and load a CLAUDE.md that has grown into a small novel. None of those feel expensive in the moment. Together they are.
The good news: you can cut a lot of waste without making the agent worse at the work you care about. It comes down to matching effort to the task, keeping the context window clean, and letting skills carry instructions only when they are needed.
What effort levels actually do

Effort controls how hard Claude thinks before it answers: how much reasoning it does and how willing it is to explore. In Claude Code you set it with /effort, and the ladder runs from low through medium and high up to xhigh and max. Higher effort means more reasoning tokens, slower turns and better results on genuinely hard problems.
The mistake is treating the top setting as 'better' in general. On a simple edit, extra reasoning mostly buys you a longer wait and a bigger bill. On a gnarly bug or a design decision, it is worth every token. Ultracode goes further still, pairing very high effort with multi-agent orchestration, and we cover it in its own guide.
- Low or medium: renames, small edits, formatting, quick questions
- High: normal feature work and debugging
- xhigh or max: hard bugs, architecture, anything you would sleep on yourself
Keep the context window clean

Every turn resends what is already in the conversation, so a session that has read twenty large files carries them all along. Use /clear when you switch to an unrelated task, and /compact when you want to keep going but shed the bulky history. Ask for summaries of long logs instead of pasting the whole thing.
Subagents help here too. When you send a subagent off to search a codebase, its file reads stay in its own context and you get back a short answer, instead of every file landing in your main session. The pstack-principle-guard-the-context-window skill teaches exactly this habit: route bulk output to files and subagents, and keep the main thread lean.
- Switch tasks:
/clear - Long but same task:
/compact - Big searches: delegate to a subagent and ask for the conclusion only
- Install:
npx @skills-hub-ai/cli install pstack-principle-guard-the-context-window
Let skills carry instructions, not CLAUDE.md
Your CLAUDE.md loads into every session, used or not. Skills work differently: until a skill is needed, Claude only sees its name and a one-line description, and the full instructions load when the task calls for them. Moving rarely used procedures, like your release checklist or migration rules, out of CLAUDE.md and into skills keeps every ordinary session lighter.
The same logic applies to MCP servers. Each connected server adds its tool definitions to the context, so disconnect the ones you are not using in a given project.
- Keep CLAUDE.md to rules you need in nearly every session
- Turn occasional procedures into skills
- Prune MCP servers per project
Skills that help you spend less
A few skills in the catalog are built specifically around cost and context. skills-tao-token-efficient-execution compiles repeated agent workflows into compact stage cards so the same job costs less each time it runs. handshake distills a long, messy session into a structured handoff, so you can start fresh without losing the thread. If you are building your own LLM app, devops-security-agent-skills-llm-cost-optimization covers model selection, batching and prompt caching on the API side.
- Install:
npx @skills-hub-ai/cli install skills-tao-token-efficient-execution - Install:
npx @skills-hub-ai/cli install handshake - Install:
npx @skills-hub-ai/cli install devops-security-agent-skills-llm-cost-optimization
A five minute checklist
Before your next long session: drop effort to medium unless the task is hard, trim CLAUDE.md to what you truly need every time, disconnect MCP servers you will not use, and plan to /clear between unrelated tasks. Then raise effort on purpose when you hit the problem that deserves it. You will spend less and, oddly, often get better answers, because the agent is not wading through yesterday's context.
Skills to try
- principle-guard-the-context-window: Apply when context is filling up: large outputs, long files, repeated reads, fan-out planning. Route bulk to subagents; keep summaries in the main thread, not …
- tao-token-efficient-execution: Token-efficient execution kit for repeated TAO agentic workflows. Compiles a skill into stage cards once (with a strong model), then executes them forever in f…
- handshake: Persistent, branch-aware context handoffs for coding agents. Handshake distills long, messy sessions into structured project state, letting Claude Code, Codex,…
- llm-cost-optimization: Reduce LLM API and infrastructure costs through model selection, prompt caching, batching, caching, quantization, and self-hosting strategies. Track spend by t…
Sources
- skills-hub.ai/blog/claude-code-ultracode-2026
- skills-hub.ai/skills/pstack-principle-guard-the-context-window
- skills-hub.ai/skills/handshake
Written by
Skills-Hub Team
Skills-Hub is the open registry for AI coding skills, with SKILL.md files synced daily from Anthropic, Google, Microsoft, and 90+ official sources. Free + MIT.