11. Context Management Like a Pro

What this section is for

Context management is the difference between “the model keeps helping” and “the model keeps getting confused.”

This section teaches you how to:

• decide what to include and exclude,
• summarize state between iterations,
• treat context like a budget,
• handle multi-file projects without losing coherence,
• prevent conflicting instructions over time.

The mental model: context is working memory

Models don’t “remember” your project. They operate on what you provide right now. If the context is:

• too small: the model guesses.
• too big: the model gets distracted and contradicts itself.
• well-curated: the model is coherent and useful.

This is why context management is a core skill—not an advanced trick.

Context principles that keep you fast

• Prefer specs over history: a stable spec block beats 50 chat messages.
• Prefer “relevant slices” over “the whole repo”: paste only what matters.
• Prefer tests and errors as evidence: they anchor behavior.
• Prefer summaries over raw dumps: compress and re-anchor regularly.
• Prefer file maps: keep a high-level index so you can navigate without flooding context.

Section 11 map (11.1–11.5)

• 11.1 What to include vs exclude from context
• 11.2 Summarize state between iterations
• 11.3 The “context budget” technique
• 11.4 Handling multi-file projects without losing coherence
• 11.5 Preventing conflicting instructions over time

Where to go next

• Start with 11.1
• Then go to 12. Safety and guardrails