Skip to content

Measured 4 of these skills across the opus-4-8 → opus-5 release — one sign flip, receipts inside #44

Description

@mavericksea-ai

Hi — I run a small OSS project (Driftproof) that re-tests agent skills across model releases: the same eval suite with and without the skill, judge sampled 5x so every score carries a confidence band, and a drift verdict only when bands separate AND the mean moves at least a 0.05 effect floor. Four skills from this repo have been in the series since Report 001; the new Report 003 covers the opus-4-8 → opus-5 boundary. Their results:

  • crafting-effective-readmes — the most interesting row in the report: a sign flip. On opus-4-8 it was net-negative (with_skill 0.760 vs baseline 0.870); on opus-5 it flipped to net-positive (0.871 vs 0.795) and IMPROVED with two band-separated cases. Anyone who benched it on the old model and dropped it would be wrong on the new one.
  • commit-work — within noise, but the baseline caught up (0.847 → 0.886, now slightly above with_skill 0.882): opus-5 does Conventional Commits discipline largely unprompted, so on that model the skill currently reads as context cost. (On the sonnet boundary it was where we caught sonnet-5 writing feat: for a bug fix — Report 001.)
  • writing-clearly-and-concisely — regressed on exactly one case, sitting exactly on our 0.05 minimum-effect floor: the weakest possible signal, flagged as such.
  • naming-analyzer — regressed on one case, driven as much by variance widening as mean drop; the skill still nets +0.068 lift on opus-5 overall.

Full tables, per-case drivers, bands, and receipts: https://driftproofhq.com/reports/003/ — earlier boundaries: https://driftproofhq.com/reports/001/ (sonnet-4-6 → sonnet-5) and https://driftproofhq.com/reports/002/ (Claude vs GPT, where two of your skills were substrate-dependent).

The suites are Apache-2.0 and grounded in each SKILL.md's own claims — every rubric criterion must trace to text the skill actually contains: https://github.com/driftproofhq/driftproof/tree/main/suites. Yours to adopt upstream if useful; npx driftproof run re-runs them on any model with your own key. And if you think any case tests beyond what a skill actually claims, I would genuinely like to know — fairness of the exams matters more to me than the verdicts.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions