down-skilling - Changelog
All notable changes to the down-skilling skill are documented in this file. The format is based on Keep a Changelog.
[1.3.1] - 2026-09-09
Other
- prompt-audit: dated prompting patterns across the skill catalogue (#791)
- Add README.md to agent-routing and down-skilling (#729)
[1.3.0] - 2026-07-15
Other
- Add agent-routing skill; update down-skilling with 2026-07-15 Haiku 4.5 calibration (#728)
[1.3.0] - 2026-07-15
Added
- "Before Distilling: Check Whether the Task Needs It" triage section.
A 2026-07-15 empirical calibration (300 Haiku 4.5 calls; evidence in the
agent-routingskill'sreferences/calibration-2026-07-15.md) measured Haiku 4.5 at 240/240 on mechanically checkable work — nested arithmetic, 30-hop chains, 25-op state tracking, trap math, 5-constraint generation — at low effort, beating Sonnet-low (17/20) on the same battery. For checkable outputs, a minimal prompt + deterministic verifier + escalate-on-fail now beats example-heavy distillation; distill fully only for judgment-shaped outputs. - Iteration warning: blind self-improvement loops measured as identity-or-drift (0/96 changes on correct answers; regression-then-freeze on the one output that did change). Loop only with an out-of-band scorer, keep argmax, stop on first regression.
Changed
- Economics section repriced to current models: Haiku 4.5 $1/$5, Opus 4.8 $5/$25 per MTok (5× both sides; was stated as ~6× on stale $0.80/$4.00 Haiku pricing).
- gaps/counting-enumeration.md: Haiku 4.5 measured 13/13 on exact word-count generation at N=10–14 under stacked constraints (Sonnet-low missed twice); decisive factor is defining the unit of counting in the prompt. Mitigations rescoped to large N and count-inside-long-output.
- gaps/multi-hop-reasoning.md: split hop types — explicit chains (lookups, state updates) measured 100% to 30 hops; the 2-3-hop caution now scoped to latent inference chains, which remain unmeasured.
[1.2.0] - 2026-05-26
Added
- add mapping-features skill for behavioral web app documentation (#432)
- restructure boot output for progressive disclosure
Other
- down-skilling: add example-calibration rules (v1.2.0) (#674)
- Remove _MAP.md files, direct agents to tree-sitting for code navigation (#545)
[1.2.0] - 2026-05-26
Added
- Source-anchoring requirement in Example Quality Criteria. Every concrete fact in an example output must trace to that example's input; invented facts cause Haiku to copy the invention pattern at runtime.
- Length-calibration requirement in Example Quality Criteria. Example output lengths must sit inside the stated output range — rules don't override the example central tendency.
- "When the input could be abstract: model the silence" subsection with a worked example showing the input → output pattern that lets Haiku acknowledge what the source omits rather than filling the gap.
- Tagged BAD/GOOD pair is now the default negative-example pattern for confabulation-prone tasks (rewriting, summarization, NL→command). Updated the distribution-table row to reflect this.
- Activation step 5: audit your example set — source-anchoring + length-calibration check before delivering the prompt. Existing Deliver step renumbered to 6.
Why
Validated by experiments at oaustegard/claude-workspace/experiments/haiku-assessment/. The un-calibrated voice-rewrite prompt produced architectural hallucination in 19/20 Haiku runs; the calibrated rerun produced 0/5.
[1.1.0] - 2026-03-02
Added
- lean harder into n-shot examples as primary steering mechanism
[1.0.0] - 2026-02-14
Other
- Update SKILL.md metadata
- Add down-skilling skill