15 Claude Mistakes That Are Quietly Costing You Tokens

Almost every "Claude is worse today" post is a context problem wearing a capability costume. The window is finite, most of what fills it was never typed by you, and the parts you did type are often working against the model rather than with it.

Nothing here is a prompt template. Every item is a thing to stop doing or a setting to change, and where a vendor has published a figure or a phrasing, it is quoted rather than paraphrased.

The three in the video are 1, 2 and 3. The rest are the ones that only show up once a session gets long or a workspace gets crowded.

1. 1
   Anthropic's own documentation, on managing tool context: a typical multiserver setup of GitHub, Slack, Sentry, Grafana and Splunk can consume about 55,000 tokens in tool definitions before Claude does any work. That is spent before your first word.
2. 2
   The same page puts the fix at over 85 percent: loading tools on demand pulls in the 3 to 5 the request actually needs rather than the entire catalogue.
3. 3
   And it is not only a cost problem. Anthropic report that tool selection accuracy degrades once you exceed 30 to 50 available tools, so a crowded toolbox makes the answers worse as well as more expensive.

Hold those three numbers in mind for the rest of the list. Most of what follows is the same shape: something is in the window that is not the job.

The paragraph at the top of your prompt that assigns a persona is the single most copied habit in prompting, and on current models it is mostly dead weight.

1. 1
   Anthropic's prompt engineering guidance, verbatim: role prompting defines expert personas and perspectives in how you phrase your query, and while this can be effective, modern models are sophisticated enough that heavy handed role prompting is often unnecessary.

Worse than neutral, sometimes. A very narrow persona, the kind that specifies a world renowned expert who never makes mistakes, constrains the answer rather than improving it. Anthropic's suggested replacement is to be explicit about the perspective you want instead of assigning a character: analyse this portfolio focusing on risk tolerance and long term growth, rather than you are a veteran portfolio manager.

• Do this instead: spend those tokens on where the sources are, and on a final instruction to check its own work before it finishes.

Both of those are things the model cannot infer. The costume is something it never needed.

Negative instructions feel precise and behave badly. To avoid a thing, the model has to hold the thing in mind, which is the opposite of where you want its attention.

1. 1
   Anthropic's guidance is one line: tell the AI what TO do instead of what NOT to do. Their own worked example is "Do not use markdown in your response" becoming "Your response should be composed of smoothly flowing prose paragraphs".

The structural point underneath: a positive instruction has a direction and a negative one does not. "Do not include numbers" leaves an infinite space of acceptable answers. "Write it as continuous prose with the figures spelled out in words" leaves one.

• Rewrite rule: for every "don't", ask what you would accept instead, and write that.

This is the setting almost nobody has opened.

• Where: click the plus button at the lower left of the chat, or type a forward slash. Hover Connectors, then Tool access.
• The three modes are Auto, which is the shipped default and lets Claude decide what to load, Always available, which loads every connector at the start of every conversation, and On demand, which loads nothing until Claude searches for what the request needs. In the product the on demand option reads "Load tools when needed".

1. 1
   Anthropic's own recommendation: Always available suits fewer than 10 connectors you use constantly. Past roughly 10, or any time conversations start hitting length limits, move to on demand.

An honest note, because it matters. The default is Auto, not Always available, so this is not silently costing everybody the full 55,000 tokens on every message. What it is, is the thing that quietly degrades every time you connect one more app and never revisit the setting. Users running around eleven connectors have reported requests exceeding the context limit by more than 33,000 tokens before the first message.

A model that has to guess where the truth lives will guess. Naming the file, the folder, the URL or the tool converts a research problem into a retrieval problem, and retrieval is the part it is reliably good at.

• Weak: use the documentation to answer this.
• Strong: the answer is in docs/api/auth.md and in the CHANGELOG entries after v4.2. Read both before answering, and quote the line you used.

The single cheapest quality instruction available, and it is one clause at the end of a prompt. Verification is a separate pass from generation, and models do not run it unless you say so.

• Add: before you finish, re read your answer against the sources and list anything you could not verify.

Related, and from the same Anthropic guidance: give the model explicit permission to express uncertainty rather than guessing. That reduces hallucination more reliably than telling it not to hallucinate, which is itself a mistake number 2.