The catalog
Claude failures are easier to manage when they are treated as operational categories. The table below is intentionally practical: symptom, likely cause, and the first repair to try.
| Failure mode | Symptom | First repair |
|---|---|---|
| Hallucination | A confident answer contains unsupported facts, fake quotes, or invented links. | Require primary sources, isolate claims, and verify against originals. |
| Refusal mismatch | Claude refuses a benign request or complies with a request that should be constrained. | Clarify authorized context, remove ambiguous harmful framing, and log refusal categories. |
| Context decay | Later turns contradict earlier instructions or ignore old evidence. | Summarize durable facts, prune irrelevant turns, and reattach source-of-truth material. |
| Prompt injection | Claude follows hostile instructions from a webpage, image, email, or file. | Sandbox tools, restrict data access, and treat external content as untrusted input. |
| Truncation | The answer ends early, misses sections, or stops mid-structure. | Inspect stop reasons, raise output budget where appropriate, or split the task. |
| False tool confidence | Claude claims it sent an email, created a file, or checked a source without evidence. | Use tool-result verification and require artifact links or command output. |
The root-cause test
Before rewriting the prompt, ask three questions. Did Claude have the right evidence? Did the surrounding product allow the required action? Did the response stop because of a documented stop reason? If the answer is no, a clever prompt is usually the wrong repair.
The Stop Reasons docs are especially useful for developers because they separate ordinary completion from truncation, refusal, tool use, and context-window overflow. That makes failures observable instead of vibes-based.
Instrument the boundary
For production systems, log source IDs, model IDs, stop reasons, token use, tool calls, and human override decisions. For editorial workflows, save the source bundle and the final human-edited claim set. For agent workflows, record which files and external pages were visible to Claude.
Failure modes get worse when Claude is treated as a single opaque box. They get easier when the workflow makes the boundary visible: what the model saw, what tools it had, what policy or product surface applied, and what evidence the final answer relies on.
FAQ
What is the most dangerous Claude failure mode?
For most users, the dangerous failure is not a visible refusal. It is a plausible answer that sounds authoritative while being unsupported or stale.
What should I log when debugging Claude output?
Model ID, prompt version, source bundle, tool calls, stop reason, token use, and any human correction are the minimum useful fields.
Primary sources
- Claude is providing incorrect or misleading responses. What's going on? Claude Help Center - March 16, 2026
Plainly states that Claude can hallucinate, produce misleading output, and should not be treated as a single source of truth.
- Context windows Claude Platform Docs - Accessed July 6, 2026
Explains what counts toward context, current context sizes, and why long conversations still need token management.
- Computer use tool Claude Platform Docs - Accessed July 6, 2026
Warns that Claude can follow instructions embedded in webpages or images, a core prompt-injection risk for agentic workflows.
- Stop reasons and fallback Claude Platform Docs - Accessed July 6, 2026
Documents API stop reasons including refusals, truncation, tool use, and context-window overflow.
- System Prompts Claude Platform Docs - Accessed July 6, 2026
Shows that Claude web and mobile system prompts are published and periodically updated, while API behavior differs.