| name | context-eng |
| description | Context engineering — memory navigation, textual graph traversal, progressive disclosure, recovery. Load when assembling context for a task or when lost. |
Skill: Context Engineering
Thin skill — the weight lives in the protocols. This is the map.
The rules that matter
- Budget per phase —
core/protocols/context-budget.md: cumulative while exploring, substitutive at EXECUTE. Never carry exploration debris into implementation.
- Memory discipline —
core/protocols/memory-sync.md: read before acting, write after acting, never touch memory-long without an ADR.
- Graph-first navigation — codemap (~1–2k tok) is always cheap: read it FIRST, then follow anchors/edges into the relevant subgraph (
graph.json). Grep an anchor (ADR-010, FEAT-2026-…) instead of reading files hunting for context. The map is falsifiable: a real module missing from it is a finding — fall back to search gracefully, then surface the gap in your report with a suggested entry (core/protocols/memory-sync.md §ritual step 2); never just grep past it.
- Progressive disclosure — skill frontmatter → skill body → hub resources (
.berserqir/skills-resources/). Load deeper only when the task demands it.
Loading order for any task
memory-short (state) → active instincts (learned reflexes — ≥ 0.7, cap 6, core/protocols/instincts.md) → codemap (map) → memory-medium §feature (sprint) → memory-long §relevant (constraints) → subgraph/files (targets) → skill + demos (method).
Recovery protocol (when lost)
Signals: repeated searches · contradicting an ADR · re-reading the same file 3× · suggesting something memory says was revoked.
Reload in order: memory-short → memory-medium → memory-long → PRD → SPECS → codemap. Then re-state the task in one line before resuming. If still lost → escalate, don't thrash.
Compression
Size budgets are enforced by hook (memory-validate). When blocked: /compress (core/prompts/compress.prompt.md). Compress at breakpoints, never mid-implementation.
