The Lean Wolf

The orchestrator ran its pre-flight patrol: worktree health check (clean), orphan PR scan, and project board query. Found 3 open PRs (#220, #222, #223) all awaiting Loki QA — active chains, not stale. Identified 6 items in "Up Next" with 3 bugs, 2 enhancements, and 1 unlabeled feature.

Priority selection began — fetching labels for all 6 candidates — when Odin interrupted with a different question.

Measured every file that loads into Claude Code's context window. The verdict was damning: fire-next-up/SKILL.md alone consumed 55 KB (1,352 lines) — 82% of all loaded skill context. It contained 5 full agent prompt templates, each repeating the same 18-line sandbox preamble verbatim.

Ranked all 13 skill files by size. The top offender was clear.

Dissected the monolithic SKILL.md and identified 8 extraction candidates — blocks that were either repeated verbatim across templates or only needed conditionally (specific flag, specific agent type).

The sandbox preamble alone was duplicated 5 times. The resume flow (215 lines) was only needed for --resume. Agent prompts were only needed when spawning that specific agent.

Created 8 template files under templates/ in parallel, then rewrote SKILL.md to reference them on demand. The sandbox preamble was extracted once and composed into prompts via {{SANDBOX_PREAMBLE}} placeholder. Each agent template is now read only when the orchestrator spawns that specific agent.

The always-loaded SKILL.md dropped from 1,352 lines (55 KB) to 316 lines (10.7 KB) — a 77% reduction in context consumed per invocation.

Branched from main as chore/extract-fire-next-up-templates, committed 9 files (721 insertions, 1,163 deletions), pushed, and opened PR #225. The diff tells the whole story: the monolith shattered into composable shards.