context-engineeringlisted
Install: claude install-skill RaveenDi/Skilloadout
# Context Engineering
Context engineering is deciding, for every single turn, exactly what the model sees: which
instructions, which examples, which retrieved facts, which prior tool results, which conversation
history — and in what order, at what level of detail, and at what point they expire. The context
window is a scarce, expensive, and *attention-dilating* resource: every irrelevant token pushes
down the salience of the relevant ones, even well under the hard token limit.
## 1. The core tradeoff: load vs. fetch
- **Pre-load** what's small, stable, and almost always needed (system instructions, the current
file being edited, a short task spec). Pre-loading costs tokens on every turn but avoids a
round-trip.
- **Fetch on demand** what's large, task-specific, or rarely needed (a codebase's full file list,
a knowledge base, old conversation history). Give the model a tool to retrieve it instead of
always carrying it — this is why agentic search often beats dumping everything upfront.
- Re-evaluate this split as a task grows: something that was "always needed" in a short task often
becomes "fetch on demand" once the context fills up with newer, more relevant information.
## 2. Keep context high-signal
- **Summarize instead of accumulating.** Raw tool output (logs, search results, file contents)
should be trimmed or summarized before it re-enters context on the next turn — the model needed
it once to decide what to do next, not forever.
- **Deduplicate.** Th