MCPs and integrations are solving the context quantity problem. You can now pipe virtually anything into an LLM - your codebase, your docs, your Slack history, your entire knowledge base. The pipes are built and the integrations exist.

What now?

Well, we’ve been optimizing for the wrong thing. The assumption has been more context, better results. Bigger context windows, smarter AI. Stuff it all in, let the model figure it out.

The research says otherwise.

More context actively hurts performance

Stanford researchers found that when GPT-3.5-Turbo had to access information buried in the middle of its context, it performed worse than when it had no documents at all.

Google DeepMind found that adding a single irrelevant sentence to math problems dropped accuracy from 95% to 72%.

Another 2025 study found that context length degrades performance even when retrieval is perfect - the sheer size of the input hurts output quality, independent of what’s in it.

But irrelevance isn’t the only problem

The research focuses on irrelevant context. That’s the easy case. Harder: context that’s relevant but wrong.

A product manager posts an overly optimistic analysis of an experiment — the conclusion gets repeated, but the flawed methodology never gets scrutinized. Someone announces a decision in Slack, but the actual direction gets changed in a meeting that wasn’t documented. A customer learning from two years ago still drives roadmap priorities, even though the market has shifted and no one’s validated whether it still holds.

Feed that to your LLM, and you’re not getting intelligence - you’re scaling mistakes with confidence.

Contradictory context might be worse than no context

Most organizations are messy and not perfectly aligned. This consequently means that the available company context is incoherent. They have a strategy deck from Q1 that contradicts the one from Q3, and no one’s reconciled them - or even noticed. Three teams building toward goals that subtly conflict, each operating on assumptions that have never been surfaced.

When an LLM encounters contradictory information about what the company believes, what was decided, or why - it doesn’t magically resolve the conflict. It either picks one arbitrarily, or hedges into uselessness, or confidently synthesizes nonsense.

Contradictions in your context become contradictions in your outputs.

Finding the smallest possible set of high-signal tokens

Anthropic’s guidance is explicit: “Good context engineering means finding the smallest possible set of high-signal tokens that maximize the likelihood of some desired outcome.”

Smallest possible, not largest available.

Andrej Karpathy frames the context window as RAM - precious working memory that must be carefully managed. You wouldn’t dump every file on your computer into RAM. Why would you dump everything into your LLM’s context?

Context quality has several dimensions

What does “high-signal tokens” mean? ISO 25012 - the international standard for data quality - defines fifteen characteristics. Seven matter for LLM context:

  • Accuracy — does this correctly represent what actually happened or was decided?
  • Completeness — is this the full picture, or are critical details missing?
  • Consistency — does this contradict other context you’re including?
  • Credibility — is this from a trustworthy source? Was the methodology sound?
  • Currentness — is this still true, or has it been superseded?
  • Traceability — can you track where this came from and how it’s changed?
  • Understandability — can the model (and your team) actually interpret this correctly?

Most teams have started thinking about accuracy. Almost no one is systematically thinking about everything else.

The skill that matters now

This is why context quality is the new bottleneck. Not context access. Not context quantity. Quality.

The companies that win with AI won’t be the ones with the most integrations or the biggest context windows. They’ll be the ones who figured out what to include, what to exclude, what’s still true, and what contradicts what.

That’s context curation. And it’s a skill most teams haven’t built yet.