Blog

23 Claude Code mistakes, checked against the docs

· Tools

This article started with a video: 23 Claude Mistakes that Devour Your Usage Limits by Jay E from RoboNuggets. It is a good list, and it made me curious how much of it Anthropic itself backs up.

So I went through the claims one by one and checked them against Anthropic's docs and articles. For claims about third-party tools and incidents, I looked for the original source.

Most of the video holds up. Below is what is confirmed, which is the correct way to work, and then the parts I don't agree with or couldn't confirm. Everything here is as of October 2026, and Claude Code changes fast.

What holds up

The mechanism behind most of it

Claude Code sends the whole conversation to the model with every message. Prompt caching makes that cheap: on Opus 5.5 and Sonnet 5.5, re-reading cached content costs 5% of the normal input price. So most of the 23 mistakes are one of two things. You carry context you don't need, or you break the cache.

Keep the context small

One chat for everything. Anthropic's best practices call this "the kitchen sink session". Unrelated tasks pile up, you pay to re-read them, and the answers get worse. The fix is /clear between unrelated tasks.

Arguing with Claude. After two failed corrections, stop correcting. The best practices page says to /clear and write a better prompt with what you learned. An Anthropic article on session management recommends rewinding: press Esc twice or run /rewind, go back to the point before things went wrong, and re-prompt. Both work for the same reason: the failed attempts leave the context.

Side questions in the main chat. Use /btw. The answer never enters the conversation history.

Follow-ups in small pieces. Every extra message re-sends the conversation and produces a new full answer. One complete prompt is cheaper and drifts less.

Open-ended exploration. The docs call this "the infinite exploration". Ask Claude to "investigate" without limits and it reads hundreds of files into your main chat. Scope the question, or give it to a subagent. A subagent works in its own context and returns only a summary. It can also run on a cheaper model with model: haiku.

A huge CLAUDE.md. The docs say to aim for under 200 lines. When the file is too long, Claude starts ignoring parts of it. Move workflow instructions into skills, which load only when needed.

Every connector always on. MCP tool definitions are deferred by default, but tool names still enter every chat. Disable servers you don't use with /mcp. Where a CLI exists (gh, aws, gcloud), prefer it: the docs call CLI tools the most context-efficient way to reach external services.

Compaction without direction. Run /compact yourself at a natural break and say what to keep, for example /compact Focus on the API changes. You can also put compact instructions in CLAUDE.md. The session management article explains why: the model is at its least intelligent point when it compacts, because the context is at its fullest, and the summary can drop details.

Don't break the cache

Switching models mid-chat. Each model has its own cache. Switch from Opus to Haiku and the whole conversation is read again with no cache hits. Anthropic's own example: 100k tokens into an Opus chat, asking Haiku a quick question costs more than letting Opus answer. Changing effort is different. On Opus 5.5, Sonnet 5.5, Haiku 5.5 and Fable 5.1 it keeps the cache.

Walking away. On a subscription the cache lives for an hour. On an API key, or once you are on usage credits, it lives for five minutes. The first message after that pays for the whole conversation again. If the task is done, /clear costs nothing. If it isn't, /compact before the break, while the cache is warm.

Pick the right model and effort

Max effort for everything. The model docs say max "may show diminishing returns and is prone to overthinking". The default on the 5.5 models is medium. The video quotes Artificial Analysis: for Opus 5.5, max scores 58 against 54 for high, at about $5.98 per task against $1.82. I confirmed the max figure on their page. The high figure I found only in a write-up that quotes them.

The biggest model for everything. The docs say Sonnet handles most coding tasks, and Opus is for complex architecture and multi-step reasoning.

Never checking usage. /usage shows which skills, subagents, plugins and MCP servers eat your plan. It also flags any behavior that takes 10% or more.

Prompt better

Vague one-liners. Say which file, which scenario, and what "done" looks like. For design work, the Opus 5.5 prompting guide lists the defaults the model falls back on: cream backgrounds, italic accent words in headlines, numbered section labels, pill-shaped buttons.

"Make it not look like AI." The same guide says a general line like this mostly swaps one default for another. Name the exact patterns you don't want.

Rules without a reason. The prompting best practices use this example: instead of "never use ellipses", say the text will be read by a text-to-speech engine that can't pronounce them. Claude generalizes from the reason.

Old prompting tricks. The Opus 5.5 guide says you can remove "think carefully" lines. In Anthropic's test, replies started sooner with no clear drop in quality. Effort is the control for thinking now. Aggressive wording like "CRITICAL: You MUST" makes newer models over-apply the rule, so the docs say to write it normally.

Skipping alignment. For a larger feature, let Claude interview you first. The best practices page gives a ready prompt: Claude asks questions, writes a spec, and you start a fresh session to build it.

Protect your machine

Turning off all safety. The docs say to use bypass mode only in isolated environments such as containers or VMs. Auto mode, where a separate classifier model reviews actions, is now the default. The story from the video is real as a report: on October 6, a developer said Claude Code in skip-permissions mode wiped his C: drive, and he restored 98% from nightly backups. A write-up links four similar GitHub issues from Windows users between September 7 and October 3. None of these is independently verified.

No backup. Checkpoints restore only edits made with Claude's edit tools. Anything done through shell commands is not covered, and the docs say checkpoints are not a replacement for git.

Installing plugins blind. The security page is direct: a plugin can run arbitrary code on your machine with your permissions. The same is true for mods, the new plugin type that changes Claude Code's own interface. Before installing a mod, claude plugin validate lists what it hooks into and what it calls.

Losing old chats. Claude Code deletes local session data after 30 days by default. The setting is cleanupPeriodDays.

What I don't agree with, or couldn't confirm

"/clear is cheaper than a new chat because CLAUDE.md doesn't load again." The docs say the opposite: CLAUDE.md is reloaded after /clear. Use /clear because it is convenient, not because it saves tokens.

Pinging the session every 30 minutes to keep the cache warm. This is not in the docs. The docs warn that a scheduled task sends your full context every time it fires. The trick also only works with the one-hour cache. I would compact before leaving instead.

"There's no real undo." Too strong. /rewind does restore file edits. It doesn't restore what shell commands did, so the advice to use git is still right.

"Haiku costs about 40 times less than Opus." True only for Haiku 5.5 on prompts up to 100K tokens. Above that the gap is 8 times.

"Blast Radius is a new mod from Anthropic." It is a sample in Anthropic's playground repository, shared without support. It holds a risky command and shows what it would change. It is not a built-in feature.

Auto mode as the answer. It is the documented default and much better than skipping permissions. But in one of the four GitHub reports, the user says auto mode approved the destructive command. Keep backups either way.

"Surprise bills trace back to one habit." The docs name two: long sessions that were never cleared, and Opus left as the default.

/btw as a side panel. In the terminal it is an overlay, and it has no tool access. It can answer from what is already in the conversation, but it can't go and check anything.

DCG as "free open source". The tool exists and blocks commands like rm -rf through a hook. Its license badge says "custom", so read the terms first.

The Jev router. Jev is a real model from TypeSafe AI, and there are community routers built on it. None of this is an Anthropic feature. There is also a tension with the cache rule above: every model switch starts a fresh cache, so check how a router handles that before trusting its savings.

Numbers I couldn't trace. The "15 cents warm, almost $12 cold" example fits Fable 5.1 list prices, where a one-hour cache write costs 80 times a cache read, but I found no source for it. The story of Fable starting ten subagents and emptying a five-hour limit in five minutes is one user's report. The author's routing results are his own numbers.

The short version

If 23 points are too many, keep two. Don't carry context you don't need. Don't break the cache. Most of the rest follows from that.

References

The video

Anthropic docs

Anthropic articles

Third-party sources