down-skilling
Distills Opus-level reasoning into explicit, procedural prompts with n-shot examples so Haiku (or Sonnet) can execute a task reliably. Use when the user says "down-skill", "distill for Haiku", or wants to delegate a task to a smaller model with high reliability.
Before distilling anything, the skill now triages: if the task's output
is mechanically checkable (schema validates, tests pass, a count is a
count), a minimal prompt plus a deterministic verifier beats a
heavy example-laden distillation — save the n-shot investment for
outputs that genuinely need judgment. That triage step, the per-gap
mitigations, and the pricing in the Economics section were all recalibrated
against measured Haiku 4.5 data on 2026-07-15 (see agent-routing's
calibration reference) —
several model-card-era priors turned out to understate Haiku, most notably
on counting/enumeration and multi-hop reasoning over explicit chains.
Full prompt architecture, the gap catalog (gaps/), worked
before/after examples (examples/), and the self-check checklist live in
SKILL.md. Version history in
CHANGELOG.md.