When Claude Misbehaves: A Fix-It Guide
Claude is useful, but it is not magic and it breaks in predictable ways. Refusals, confidently wrong answers, a chat that "forgets," and usage limits are…
Claude guide · as of July 10, 2026 · 4 minutes read · details change — confirm current specs on claude.ai
Claude is useful, but it is not magic and it breaks in predictable ways. Refusals, confidently wrong answers, a chat that "forgets," and usage limits are the four complaints that come up most. Nearly all of them have a quick fix once you know what is actually happening underneath. This guide is ordered roughly by how common the problem is.
Start with the 30-second checks
Before you assume something is broken, rule out the boring causes:
- Reload the page or restart the app. A stalled response is often just a dropped connection.
- Check whether it's just you. If claude.ai is slow or erroring for everyone, look at Anthropic's status page before troubleshooting your own account.
- Start a new chat. Long, messy threads degrade quality — a fresh chat with a clear ask often fixes "it suddenly got dumber."
- Confirm which model you're on. Speed and quality complaints are frequently just the wrong tier for the job.
Refusals: when Claude won't help
Sometimes Claude declines a request it could reasonably assist with — a "false refusal." It usually happens when a prompt pattern-matches to something risky even though your intent is fine: security research, medical questions, writing a villain, quoting a long source.
What actually works:
- State your purpose and context. "I'm a nurse charting a patient" or "this is for a novel" gives Claude what it needs to say yes.
- Rephrase without the trigger framing. Ask for the underlying thing plainly.
- Break it into smaller, obviously benign steps.
If a refusal is genuinely wrong, report it with the thumbs-down control — that's how the guardrails get tuned. Some limits are firm by design, and no rephrasing will move them.
Wrong or made-up answers
This is the failure mode to take seriously: Claude can be confidently wrong. It will state a fake citation, a wrong date, or an invented function name in the same calm tone it uses for correct answers. Two structural reasons:
- Knowledge cutoff. Claude's training has a cutoff date, so anything newer is guesswork unless it searches the web. For recent events, turn on web search or paste the source yourself.
- Hallucination. For anything that matters — legal, medical, financial, or work you'll publish — verify independently. Ask "what's your source?" and actually check it; don't trust a link just because it looks real.
Giving Claude the real document, data, or code beats asking it to recall from memory almost every time.
"It forgot what we were talking about"
Claude has a large context window — up to roughly 1M tokens on some models — but it is not infinite, and by default it does not remember across separate chats. In a very long thread, early details fall out of view.
Fixes:
- Summarize and restart. Ask for a recap, then paste it into a fresh chat.
- Use Projects to hold documents and instructions that should persist across many chats.
- Re-attach the file instead of trusting recall from 200 messages ago.
Rate limits and "you've hit your limit"
Every tier — Free, Pro, Max, Team, Enterprise — has usage limits, and they reset on a rolling window rather than a fixed daily clock. Paid plans raise the ceiling; no plan is truly unlimited.
| Symptom | Likely cause | First thing to try |
|---|---|---|
| "Out of messages" mid-task | Hit your rolling usage cap | Wait for the reset window, or switch to a lighter model |
| Everything feels slow | Peak load, or a heavy model on a simple task | Try Haiku 4.5 for quick work |
| Big file won't process | Over the context or upload limit | Split the file; summarize in pieces |
| Long agent run stops | Claude Code / API limits or cost caps | Check current limits and your usage |
Don't memorize numbers — check the current plans and limits at claude.ai / anthropic.com, since tiers change over time.
When this is the wrong tool
Troubleshooting won't help if you're asking Claude to do something it structurally can't:
- Guaranteed-current facts with no source — it has a cutoff and can be wrong; use web search or supply the source.
- Fully offline, private-by-architecture work — Claude is cloud-only. Turning off training does not make a request local. If the data truly can't leave your machine, run a local model instead.
- Image generation — Claude reads images and can build visuals via code, but it has no native painter like DALL-E.
- Always-on voice, or a phone / Siri / smart-home assistant — voice is nascent, and Claude isn't embedded in your operating system.
For those, a different tool is the fix, not a better prompt.
