Claude 5.5 Model Mistakes you may still be making (23 Tips and Fixes)

Summary

The video presents 23 ways to use Claude more effectively, focusing on reducing token and usage costs, improving output quality, and protecting files and accounts. The advice is grouped into easy, intermediate, and advanced practices. It draws on Anthropic guidance, comments attributed to Anthropic engineers, independent model tests, and power-user workflows.

Easy fixes

1. Avoid “kitchen sink” sessions. Use separate chats for distinct tasks. Unrelated conversation can distract Claude, and chat history is reread as a session continues. Use /clear when starting a new topic while retaining project instructions.

2. Don’t switch models mid-chat unnecessarily. Each model has its own prompt cache, so switching can cause the conversation to be reread without the same discount. Pick a model for the task and stick with it. Changing effort may preserve the cache on newer models, but the speaker says this feature is in beta and should be checked on your setup.

3. Rewind after repeated mistakes. Rather than repeatedly correcting a bad result, rewind to before the error—suggested after two failed corrections—using Escape twice or /rewind.

4. Keep side questions out of the main task. Use /btw for a quick question so its answer appears separately and doesn’t add a detour to the main chat history.

5. Reserve maximum effort for demanding tasks. The video says max effort can lead to overthinking and may cost more for limited gains. It cites an Artificial Analysis test in which Opus 5.5’s max effort scored about four points above high effort but cost more than three times as much per task.

6. Give specific prompts. State the intended output, audience, goal, and tone instead of issuing vague requests. For repeated work, use a “harness”—such as a design system or reusable skill—to provide standards and procedures.

7. Explain the reason behind rules. Briefly say why a constraint matters so Claude can apply it sensibly rather than following it blindly.

8. Specify what “not AI-looking” means. General requests to avoid a generic AI style may just produce another default style. Name unwanted patterns or provide a design system.

9. Don’t settle for the first draft on important work. Request multiple independent versions and have Claude evaluate them against the goal. This can help with creative tasks, though each version uses additional resources.

10. Avoid drip-feeding requirements. Put the requirements in one well-formed prompt where possible; repeated follow-ups can consume usage and cause output to drift.

11. Retire outdated prompting tricks. The speaker says newer models respond better to clear instructions than to emphatic phrases such as “think carefully” or all-caps commands. Use the effort setting to control how much reasoning is applied.

Intermediate practices

1. Account for cold caches after breaks. The video says subscription prompt caches may go cold after roughly an hour. It recommends /clear if a task is finished or /compact before leaving a long session. It also describes using a periodic /loop ping to keep a cache warm. One cited example compares about $0.15 to continue a warm, roughly 400,000-word session with nearly $12 after the cache went cold.

2. Keep CLAUDE.md concise. Treat it as a router to relevant skills, documents, and project folders, not a repository for every instruction. The speaker recommends keeping it under 200 lines, saying overly long files cost usage and can make important instructions easier to miss.

3. Don’t disable safeguards indiscriminately. The speaker warns that approving everything or skipping permissions can expose files to destructive actions. Suggested protections include auto mode, a “Destructive Command Guard” add-on, and a deletion-review mod described as showing affected files before the user proceeds.

4. Back up important files before major edits. Make dated copies or use Git, which the video describes as a way to save project snapshots and roll back changes.

5. Limit always-on integrations. MCP servers and other connectors add tools and can increase usage and access risk. Audit which connections are actually used; consider command-line interfaces (CLIs), which the speaker says can be cheaper and faster for suitable tasks.

6. Scope large investigations or delegate them. Avoid unbounded searches that pull many files into the main conversation. Narrow the target or use a sub-agent to investigate separately and return a summary; the sub-agent can use a lower-cost model for routine research.

7. Check usage with /usage. The speaker says this can break down usage by skills, sub-agents, plugins, and MCP servers. They also suggest a custom mod that displays short-term and weekly usage and flags expensive sessions.

8. Align on larger projects before building. Have Claude ask clarifying questions first, particularly for complex systems such as client portals or onboarding workflows. The speaker describes using an “align” skill to confirm requirements before implementation.

Advanced practices

1. Change the session-history retention period if needed. The video says Claude Code deletes local session history after 30 days by default. It recommends changing the cleanup period in settings—for example, to 365 days—and using /resume to return to retained sessions.

2. Review plugins and mods before installing. Plugins may run scripts on the machine and add instructions or tokens to every turn. Check what they install, their cost, and whether they are necessary; the speaker describes using a “personalize” skill to assess repositories or plugins.

3. Direct compaction deliberately. When a long chat needs summarizing, compact at a natural stopping point and specify what must be retained, such as a client brief or price list. The video also suggests adding these preferences to CLAUDE.md.

4. Route tasks to an appropriate model. Don’t use the largest model for every small task. The suggested basic approach is Sonnet for defined work and Opus for exploratory work; an automated router can choose a model per prompt. The speaker reports that their router sent nine out of ten prompts to Sonnet or Haiku over three weeks, with routing costing 13 cents, and says Haiku was about 40 times cheaper than Opus.

Guides, demonstrations, and evidence mentioned

• The speaker offers a free PDF containing the tips, resources, and prompts from the video.
• Several recommendations are illustrated with prompts shown onscreen, including checking whether effort changes preserve the cache, auditing MCP usage, and building a usage-display mod.
• The video cites Anthropic documentation and engineers for guidance on prompting, effort, caching, compaction, and session behavior.
• It references an independent Artificial Analysis comparison of Opus 5.5 effort levels, as well as examples and reports from power users. These are presented as the speaker’s sources and anecdotes, not as independently verified findings here.

Main speaker and sources

• Main speaker: Jay, introducing himself as the creator of the RoboNuggets channel.
• Sources discussed: Anthropic documentation and engineers, Artificial Analysis benchmarks, and tips or reports attributed to Claude power users.