Three context engineering mistakes that degrade AI performance

Executive overview

Bad context degrades even the best models — accuracy can drop by up to 30%. The three failure modes are context clashing (contradictory inputs), context confusion (too many tools), and context distraction (bloated windows that cause repetitive behaviour).

Each mistake has a direct fix: front-load coherent context, limit and dynamically load tools, and offload or summarise history rather than letting it accumulate.

Garbage in, garbage out still holds — and a bloated context window is a form of garbage.

Context clashing: contradictory inputs hurt accuracy

• Incrementally adding context across a long conversation introduces contradictions that degrade responses.
• Research showed accuracy drop from 98% to 64% on high-quality models with contradictory sharded input.
• In vibe coding, repeated failed attempts compound: the longer the conversation runs, the worse the fixes get.
• Fix: front-load all relevant context in a single message rather than drip-feeding it.
• For complex inputs, run a brain dump through a summarising LLM first to detect and flag contradictions before passing to the working model.
• Always audit system prompts for conflicting instructions — modern models follow instructions precisely, so any contradiction will be acted on.

Context confusion: too many tools cause wrong calls

• As the number of available tools grows, the rate of incorrect tool selection increases measurably.
• The practical ceiling for today's models is around 20 tools; beyond that, accuracy degrades noticeably.
• Similarity between tool names compounds the problem — the model struggles to distinguish near-identical options.
• Fix 1 — tool loadout: only inject the tools relevant to the current task, not the full set.
• RAG-MCP implements this: a smart retriever uses semantic search to pull 5–15 matching tools from an external database, then passes only those to the working model.
• Fix 2 — context offloading: keep long-term or infrequently needed data outside the context window and retrieve it on demand.

Context distraction: long windows cause repetitive behaviour

• Large context windows are not inherently better — Gemini 2.5 Pro begins repeating past actions once the context exceeds 100k tokens.
• The model fixates on previous steps rather than generating novel solutions, producing a loop of ineffective repetition.
• Fix: summarise previous context through a separate LLM before passing it to the working model, keeping only what is relevant.
• In vibe coding: start a fresh conversation for each new feature or unresolved bug; summarise what was achieved and carry only that forward.
• Offloading also applies here — externalise history and retrieve selectively rather than accumulating it inline.

Context offloading: shared fix across confusion and distraction

• ChatGPT memory: saves user-specific context externally and retrieves it when relevant, avoiding persistent bloat.
• Claude's think tool: routes extended reasoning to an external scratch pad; only the distilled output re-enters the main context window.
• To-do / planning documents: a macro plan lives outside the conversation; the model checks it off task by task, starting a fresh context window for each, with only a compressed summary of prior progress carried forward.