Context Clash

Canonical version: Context Clash.

Context clash is "when you accrue new information and tools in your context that conflicts with other information in the context" (Drew Breunig, "How Long Contexts Fail", June 2025). Two parts of the context say different things, and the model has to pick, often silently and not always the same way.

It's one of Breunig's four long-context failure modes, next to Context Poisoning (an error gets in and keeps being referenced), Context Distraction (the context gets so long the model over-relies on it) and Context Confusion (irrelevant content steers the answer). I use his terms across the vault: every contradiction is clash.

Evidence

A Microsoft and Salesforce study took benchmark prompts and split ("sharded") them across several turns, the way real conversations reveal a task bit by bit. Performance dropped by 39% on average. OpenAI's o3 went from 98.1 to 64.1. The early turns hold the model's first attempts and wrong assumptions, and they clash with the information that arrives later. The model doesn't recover.

The clearest public example of clash in a harness came from Anthropic. When it rebuilt Claude Code's system prompt for the Claude 5 models, its own transcripts showed instructions colliding inside one request: "leave documentation as appropriate" from one source, "DO NOT add comments" from another. The model could still work out the intent, but it spent reasoning on reconciling its own instructions before doing the work. Anthropic cut more than 80% of the prompt.

How it happens

• Rule conflicts: two instructions that contradict each other ("always be concise" next to "give detailed explanations with examples")
• Harness collisions: in an agent, most of the context is assembled for you (tool schemas, skills, MCP server instructions, subagent results, compaction summaries). Each piece was written by someone else, at a different time. Nobody reads the full window, so nobody sees the conflicts
• Multi-turn drift: early turns with partial information or wrong guesses, followed by turns that correct them
• Temporal inconsistency: the codebase or project changed, and context from different periods describes it differently
• Multi-source merging: context from different authors, tools or agents with incompatible assumptions
• Memory against instructions: AI memories that say one thing while current instructions say another, because the approach changed and the memories didn't. Over months this gets worse: old facts are rarely deleted, only contradicted (Strategic Forgetting)
• Stale documentation: CLAUDE.md files or skills that describe patterns, functions or conventions that no longer exist

Why it's dangerous

Clash produces outputs that look confident and plausible but are subtly wrong. The model rarely flags a contradiction; it resolves it, and the resolution can change from one run to the next. Users can't predict which instruction will win, and the model may invent a synthesis that wasn't in any source.

Clash is often where unchecked AI Context Rot ends up. Context accumulates without pruning, staleness turns into contradiction.

Mitigation

• Single Source of Truth (SSOT): one canonical place for each piece of context. Say a rule in the tool description, not in the description AND the system prompt AND a skill
• Clear precedence: decide what wins (user > project > skill > default) and write it down (Context Inheritance)
• Prune, don't patch: remove outdated context rather than adding corrections on top of it (Context Hygiene)
• Restart instead of correcting: if a conversation started with wrong assumptions, a fresh session with the full, correct request often beats more turns of corrections
• Tool loadout: fewer tools and skills means fewer chances to collide (Progressive Disclosure)
• Versioned context: Context-as-Code makes changes visible and reviewable
• Read the assembled context once in a while. Most harnesses can dump it, and it's usually a surprise
• Conflict detection: ask an AI to scan your context files for contradictions before they cause trouble

References

• Drew Breunig, "How Long Contexts Fail" (2025-06-22): https://www.dbreunig.com/2025/06/22/how-contexts-fail-and-how-to-fix-them.html
• Drew Breunig, "The State of Context Engineering" (2026-10-24): https://www.dbreunig.com/2026/10/24/the-state-of-context-engineering.html
• Laban et al., "LLMs Get Lost In Multi-Turn Conversation" (Microsoft, Salesforce, 2025): https://arxiv.org/abs/2505.06120

Related

• Context Engineering
• Context Confusion
• Context Poisoning
• Context Distraction
• Context Bloat
• Harness Engineering
• Progressive Disclosure
• AI Agent Skills
• Fighting the Weights
• AI Context Rot
• Single Source of Truth (SSOT)
• Context Inheritance
• Context Hygiene
• Context-as-Code
• Strategic Forgetting
• AI Instruction Drift

About Sébastien

Ready to get to the next level?