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.
Skill folder: down-skilling
Release of down-skilling version 1.3.1
📥 Download & Install
To install:
- Click the download link above (ignore the "Source code" archives below - they're auto-generated by GitHub)
- Go to Claude.ai Skills Settings
- Upload the downloaded ZIP file
- Requires paid Claude Pro or Team account
See official documentation for more details.
Recent Changes
db4e3e1 prompt-audit: dated prompting patterns across the skill catalogue (#791)
b2c5feb Add README.md to agent-routing and down-skilling (#729)
15d4e9b docs: Update CHANGELOG.md for released skills