How to avoid Claude 5.5 mistakes and save tokens

TL;DR

The key is to split tasks into separate chat sessions and pick a single model per task to avoid costly token rereads. Use clear instructions with reasons, harness prompts for repeated tasks, and avoid long back and forth on a single thread. Disable unnecessary side questions and rely on side panels for quick checks.

Transcript

With the recent release of the cloud 5.5 models, we are now in a whole new era of using cloud. Opus and sonet 5.5 now think differently. We now have the capability to create our own mods for cloud code. And even the way the cash saves you money has changed. So several of the habits that worked before are now quietly eating your usage limits and may... Read More

Key Insights

• A kitchen sink session keeps one chat open for many tasks and can cause Claude to reread the entire conversation, increasing token usage.
• The prompt cache in Claude 5.5 reduces the reread cost but still adds usage and can degrade answers with unrelated history in the same session.
• Switching models mid chat triggers full rereads because each model has its own prompt cache, so start with a chosen model for a task.
• Using a quick slash command like /clear can reset prompts so you don t reload past context for new topics.
• Max effort can yield diminishing returns and cost more than high or medium settings, so reserve max for big jobs and use medium or high otherwise.
• One sentence prompts without context lead Claude to default styles; provide context such as audience, goal, and definition of done to improve results.
• Asking for a single first draft is risky; request multiple independent versions and evaluate them to pick the best fit for your goals.
• Drip feeding follow ups rereads the whole chat each time, so write a complete prompt in one go or rewind to a good point before continuing.

Questions & Answers

Q: How should you structure Claude prompts to avoid vague results?

Provide context beyond a simple instruction, specify the audience, the goal, and what done looks like. If possible, create a harness or design system that Claude can reference for repeat tasks. This reduces guesswork and aligns outputs with your standards from the start.

Q: What is the risk of keeping one chat session active for many tasks?

Keeping a single session active causes Claude to reread the entire conversation for each new prompt, which increases token usage and can distract the model, leading to worse answers over time. Start a new session for each distinct task to minimize this risk.

Q: Why should you avoid switching models mid chat in Claude 5.5?

Each model has its own prompt cache, so switching causes the conversation to be reread with less discount, costing more tokens. Choose a model at the start of a task and stay with it unless a clear benefit justifies the switch.

Q: What is the purpose of the /clear command in Claude chats?

The /clear command helps reset system prompts and skills loadouts so you can start fresh on a new topic without carrying over irrelevant context, reducing token usage and preventing cross-topic contamination in your task.

Q: When is max effort worth using in Claude 5.5?

Max effort can show diminishing returns and may overthink, costing significantly more tokens. Use it for rare high-stakes tasks, but for normal work, medium or high are typically sufficient and more cost effective.

Q: How can you prevent side questions from bloating the main chat history?

Use a side prompt approach, such as a /btw command, to get quick answers without inserting them into the main chat history. This keeps the main task clean and avoids reprocessing side questions on every response.

Q: What should you do instead of taking the first draft as the final output?

Ask for multiple independent versions and have Claude or your agent evaluate them against your goals. This burst of ideas can yield better results for creative work, though it costs more tokens, so use it selectively.

Q: Why is giving Claude rules without a reason problematic, and how should you fix it?

Rules without reasons can cause the model to follow guidance blindly and miss the intended goal. Always include a short reason for each rule, so Claude understands the context and can apply the rule appropriately rather than mechanically.

Summary & Key Takeaways

• Claude 5.5 users should avoid keeping a single kitchen sink session for multiple tasks to prevent distraction and extra token rereads, as batch memory increases costs and reduces accuracy.

• Switching models mid chat can erase prompt caches and waste tokens, so pick a model at the start of a task and stick with it unless new model behavior allows cache retention.

• Use practical prompting practices like providing reasons for rules, naming exact design patterns, asking for multiple drafts, and using side prompts to keep main tasks clean and efficient.