CodexGuild Knowledge Base
Context engineering: the discipline that replaced prompt engineering
Canonical as of Aug 20, 2026
Context engineering: the discipline that replaced prompt engineering
2025-2026 industry consensus: stop micro-tuning prompts, start engineering context — what's in the window (retrieval, tools, memory), compaction, and caching. System prompts matter less than context composition.
Context engineering (2025-2026)
As of: 2026-08
The shift
Prompt engineering (clever single-message phrasing) gave way to context engineering: systematically deciding what occupies the model's attention window —
- Retrieval: the right chunks, not everything (RAG entry).
- Tool schemas: minimal, well-described; tool descriptions are instructions (and attack surface — see MCP entry).
- Memory: durable facts vs session state; compression when history grows.
- Compaction: summarizing older turns to fit the window without losing task state (agents losing the plot mid-task = compaction bug).
- Caching: prompt-cache-friendly layouts (stable prefix, volatile suffix) — 5-10x cost cut on agentic loops.
Practices that proved out
- Stable system/tool prefix, dynamic data last — cache hits fund everything.
- Summarize-then-continue beats truncation on long tasks.
- Explicit "as of" dates on knowledge (the CodexGuild KB model) — models reason about freshness when given dates.
- Eval changes to context like code changes: golden tasks, before/after.
The job title changed; the skill is information architecture for attention.