Context Engineering

(my viewer's notes on Advanced Context Engineering for Agents)

Optimize the context window for

• correctness
• completeness
• size
• trajectory

<-- note that "size" is 3rd! -->

Bad context data attributes!:

• incorrect
• missing
• noisy

Output quality is inversly proportional to context window size: Using more context does not yield better outcomes!!

implement "intentional compacting" any way you can (automated compacting is still an anti-feature at this point)

Goal!!

Keep context use <40%

Three phase approach

• research

  • understand the current system (how it works)
  • find all relevant files

  (to a first approximation, these are == /init in Claude Code)

  • explore causes of $foo bug

• plan

  • outline exact implementation steps
  • include filenames/line #s/snippets
  • be very explicit about how we will test and verify

• implement

  • go write the code duh

  "If properly planned, the implementation is easy and expected."

  (keep context < 40%!)

There's a near-exponential explosion here

                   1 bad LOC == a bug / 1 bad LOC
          1 bad Line Of Plan == 10--100 bad LOCs
      1 bad Line Of Research == 10--100 bad LOPs == 1,000+ bad LOCs
          1 bad Line Of Spec == etc...           == 10,000+ bad LOCs
       1 bad Line Of Command == etc...           == 1000,000+ bad LOCs ("Command" referring to markdown/CLAUDE.md/etc..)