A large context window sounds like a dream: you throw in documentation, the entire codebase, past decisions, logs, and conversation, and the model "has everything." In practice, "everything" often means a lot of irrelevant, outdated, or contradictory context.
After enough AI coding sessions, it becomes clear that the problem isn't just context capacity. The problem is context quality, the timing of when you introduce it, and a clear decision about what the model actually needs to do.
More context can reduce focus
The model doesn't always know what's most important. If you give it ten possible directions, it will often try to satisfy all of them. This can produce an implementation that looks diligent but solves the wrong problem or pulls decisions from documentation that's no longer valid.
Bad signal in context
- old architectural decisions without a note that they're outdated
- unfinished experiments that look like standards
- too many tool definitions that aren't relevant to the task
- a long conversation history full of direction changes
Good context is selective
For most tasks, a smaller, well-chosen package is better: relevant files, current specification, constraints, expected output, and a few examples of local style. If the model needs more, it can ask for it or read targeted files.
This is a similar principle to software architecture: boundaries exist to reduce the number of things you need to hold in your head at once. An AI workflow without boundaries quickly becomes slow, expensive, and unpredictable.
Specification beats a pile of history
If the task isn't clearly defined, the model will fill in the gaps. More context then just provides more material for wrong guesses. A short specification with the goal, non-goals, edge cases, and acceptance criteria is often worth more than thousands of lines of conversation.
That's why AI-assisted development works better when combined with a spec-driven approach. The specification gives the model intent, and selective context gives it enough local knowledge for implementation.
A practical checklist before a big AI session
- What is the one decision or change I want to get?
- Which files are actually relevant?
- Which documents are outdated or contradictory?
- What output do I expect: patch, plan, review, or research?
- Which risks must be checked before finishing?
Conclusion
A large context window is useful, but it's not a strategy. Strategy is knowing what the model needs to do, which information it needs, and which information gets in the way.
Teams that want to use AI tools seriously need to think about context management as part of the engineering process: specifications, relevant sources, clear boundaries, and review. Without that, more context just increases the surface area for wrong conclusions.
Josip Budalić
HOTFIX team
Josip runs HOTFIX d.o.o. and works on software architecture, AI-assisted development workflows, codebase modernization, and practical software delivery.