Two High-Stakes Symptoms: Hallucination and Context Crowding
CoreDiagnose underperforming prompts and outputs · Difficulty 2/5
Explanation
Fabricated Facts and Citations
When Claude states wrong facts or invents citations with apparent confidence, that is Hallucination, not a phrasing or tone problem. The targeted fix is to ground the answer in source material, require verifiable citations, and explicitly allow the model to say "I don't know" rather than fabricate an answer to satisfy the request. Fixes that do not touch grounding -- asking for more confidence, shortening the prompt, raising temperature -- leave the actual cause untouched.
Quality Degrading Late in a Long Chat
A distinct symptom is output quality holding up early in a conversation and then degrading only near the end. This pattern points to a crowded Context Window: as a chat accumulates turns, low-signal content crowds out what matters and the model effectively loses the thread. The fix is to manage the context directly -- summarize the conversation, start a fresh chat, or persist the key instructions and facts in a Project so they don't have to survive inside an ever-growing transcript.
Why These Two Are Worth Separating
Both symptoms are easy to mis-diagnose as "the model got worse" or "try a different platform." In fact both have a specific, recognizable trigger (missing grounding; a crowded window) and a specific fix that has nothing to do with model choice or prompt wording in general.
Common exam traps
- Treating a fabricated citation as something temperature, tone, or brevity instructions can fix -- only grounding and a verifiable-citation requirement target Hallucination directly.
- Responding to late-conversation degradation by adding more prohibitions to the prompt or switching to a smaller model -- neither addresses a Context Window problem; summarizing, restarting, or persisting to a Project does.
Key Takeaways
- Hallucination (confidently wrong facts or invented citations) is fixed by grounding in source and requiring verifiable citations, not by tone or temperature changes
- Allowing Claude to say 'I don't know' is part of the hallucination fix, not a separate concession
- Quality that only drops late in a long chat signals a crowded context window, not a weaker model or a worse prompt
- The context-window fix is to summarize, restart, or persist key info in a Project -- not to add more prohibitions or downsize the model
Glossary Terms
Attention degradation caused by a context window filling with irrelevant, stale, or low-signal content, even when technically there is still room left in the window. Distinct from running out of space (a hard context-window limit) and from position effects (attention bias by location within the window) -- context rot is specifically about signal-to-noise degrading as low-value tokens accumulate.
The maximum amount of text (measured in tokens) that Claude can process in a single request. Includes both input tokens (prompt, history, tool results) and output tokens. Exceeding the context window causes an error or requires context management strategies.
Confident, plausible-looking content that Claude generates which is false or fabricated -- an invented statistic, citation, source, or quote. Delivered with the same fluent tone as accurate content, which is what makes it hard to detect by tone alone. Concentrates in specific-looking details, at the edge of the model's knowledge, and inside long outputs.
Related Concepts