| name | context-window-engineering |
| description | Decide what actually goes in the context window — static instructions vs just-in-time retrieval vs conversation history vs tools — with a per-section token budget and lost-in-the-middle-aware ordering. Reach for this when the window is bloating, quality is dropping as context grows, cost/latency is climbing, or you're unsure what to keep vs compress. Pairs with prompt-pattern-selection. |
Skill: Context-window engineering
The window is a budget, not a scratchpad. This skill decides what earns its
tokens and in what order.
Step 0 — One opinion up front
More context is not free, and past a point it hurts. Cost and latency scale
with fill, and accuracy can drop (the lost-in-the-middle effect + dilution). If
quality falls as you add context, remove context.
Step 1 — Classify every candidate
For each thing you're tempted to include, classify it (trace
../../knowledge/prompt-decision-trees.md §3):
- Needed every call → static system context (instructions, schema, invariants,
few-shot).
- Query-dependent knowledge → retrieve just-in-time (
ai-rag-engineering),
budgeted.
- Conversation history → keep recent verbatim; summarize/compress older turns.
- Everything else → probably doesn't belong.
Step 2 — Budget tokens per section
Assign a token ceiling to each section (system, retrieved, history, tools, output
headroom). The sum must leave room for the response. Write the budget down.
Step 3 — Order for the model, and for the cache
- Lost-in-the-middle: put the decisive material at the start or end, not
buried in the middle.
- Prompt caching: put stable content first (system, schema, few-shot) so it
caches; put variable content last. This is a cost and latency lever.
Step 4 — Define eviction order
When the window approaches its limit, what gets dropped or compressed first?
(Usually: oldest history → least-relevant retrieved chunks → verbose examples.)
Make it a rule, not an accident.
Step 5 — Hand off
- The retrieval that fills a query-dependent slot →
ai-rag-engineering.
- The model's window size / caching behavior →
ai-coding-model-guidance /
claude-api.
- Measuring whether a context change helped →
prompt-reliability-engineer.
Output
A context-inclusion plan: each section classified, a per-section token budget,
the ordering (lost-in-the-middle + cache-aware), and the eviction order.
