Maximizing the value of your Claude Code sessions

TL;DR Highlight

Using /clear, @-mentions and cache-aware habits makes Claude Code finish the same work with fewer tokens.

Who Should Read

A developer who uses Claude Code daily and burns through token limits or costs too quickly.

Core Mechanics

Three things set a token's price
- Model, input vs output, and cached or not decide the price.
- Output is priced at roughly 5x input.
- Many output tokens are thinking tokens, and /effort controls how many.
Cache reads cost 0.1x input price
- If a request starts with the same tokens as before, the server reuses the stored state (prompt caching).
- Writing to the cache costs up to 2x normal input.
- A write happens once per token, while the 0.1x reads happen on every later turn.
Switching model or effort busts the cache
- /model and /effort are part of the cache key, so changing mid-conversation re-prefills the whole conversation.
- Fast mode is also part of the key, and the re-prefill is billed at fast mode prices.
- The cache expires after an hour on a subscription or five minutes on an API key (ENABLE_PROMPT_CACHING_1H=1 makes it an hour).
Everything in context is resent every turn
- Files read and command outputs are sent again on every later turn.

Community Reactions

A /handoff skill as an alternative
- Commenters say it saves key session context to a short file and starts a fresh session with /continue.
- One reports better results doing this every ~20 messages than running long sessions.
Question why the cache is tied to effort
- One commenter asks whether effort could be a decoding-only setting.
- They say follow-up questions after hard analysis need less effort, but switching is costly.
Pushback that @-mention is an antipattern
- One view is that large files get fully loaded, while Read can target only the relevant part.
- Another reports @ search in the desktop app returns worse results than the CLI.
Unexplained cache write spikes
- One user with a 1h TTL and no model or effort changes saw cache writes grow from 400K to 800K and 2M.
- They noticed it around /btw and code review requests.
Criticism that this shifts burden to users

How to Apply

Check /context, /model and /effort today
- Open a fresh session and run /context to see what CLAUDE.md and MCP tool definitions take up.
- Turn off unused MCP servers via /mcp and confirm /model and /effort are deliberate.
/clear per task and /rename if you'll return
- Run /clear before starting new work, and /rename first if you may want the session back.
- When the earlier part of a task is done, run /compact and say what to keep.
Put quiet commands in CLAUDE.md
- Add your 2-3 daily commands with quiet flags (e.g. npx vitest run <file> --reporter=dot).
- For big-output jobs like log review, tell Claude to do it in a subagent.
/compact before breaks, switch models at session start
- If you'll be away over an hour, run /compact while the cache is still warm.
- Change model, effort or Fast mode at session start or right after /clear.

Try It

copy & run

# Paste into a fresh Claude Code session (a simple prompt version of the article's tips)

/context

Based on the result above, answer in this format.
1. Items loaded at startup that look unnecessary for this task (CLAUDE.md sections, MCP servers)
2. Workflow-specific instructions in CLAUDE.md that should move into skills
3. A draft CLAUDE.md snippet for the commands I use all day {{2-3 frequent commands}}, with quiet flags

Next: turn off unneeded servers with /mcp, and run /clear before your next task.

Terminology

prefill / decodePrefill is the phase where the model reads the whole request at once, and decode is where it writes output one token at a time. Think skimming a page versus writing it by hand letter by letter.

Prompt cacheA feature that reuses computation when the start of a request matches an earlier one. It's like saving earlier work and only computing the new tail.

SubagentA helper agent with its own context window. It doesn't see your conversation and returns only its final answer.

/compactA command that replaces a long conversation with a summary. Because the conversation is rewritten, the old cache no longer matches.

/rewindA command that cuts the last few turns off the end. The earlier cache stays intact, so it costs nothing.