Replace the overlapping human-writing workflows with one automatically selected explain-clearly skill for plain English, documentation, messages, and HTML explainers. It combines meaning-preserving editing, explanation structure, typography, accessibility, and verification, with supporting references by output type.

The installer selects the new skill and archives technical-writing, simple-english, and clarity outside discovery after a successful upgrade. Local customisations remain recoverable. Repository guidance and agents now point to the consolidated skill. The changeset marks the retired skill name as a breaking change.

The existing promptfoo harness now has a three-arm artifact suite, deterministic grader proofs, lifecycle isolation controls, and repeated model-rubric calibration. The suite preserves generated content and prevents deferred cleanup or timed-out cases from contaminating later arms.

Evidence

• Automatic selection: 12/12 writing checks across three repeats; 6/6 unrelated-request controls pass.
• 18 offline grader checks, 3 lifecycle controls, and 10 repeated model calibration cases pass.
• Revised prose: forced skill 14/15 metrics across three repeats; automatic selection 13/15. Unsupported reassurance remains a real miss. One automatic-arm failure was a rubric false positive on purpose versus outcome; the rubric was corrected and calibrated, without retrospectively changing that recorded score.
• HTML snapshots were inspected in Chromium at 1280px and 320px. Keyboard input, presets, and 59/60/89/90-second transitions work. A later snapshot fixes the initially misplaced 60/90 labels and the conditional promise. Its endpoint label still overflows by 9px at 320px; guidance now covers inward anchoring, with confirmation pending.
• Frontmatter, eval wiring, installer migration, and focused repository checks pass. git diff --check passes.
• The full repository test runner stops at a pre-existing architecture-guidance assertion about the unchanged twelve-factor skill. The missing phrase and source are unchanged from the base revision.

Why this remains a draft

Prose faithfulness is not yet reliable across repeats. HTML agent runs timed out after 15 and 7 minutes, so the full automated three-arm HTML comparison is incomplete; browser findings from saved snapshots are separate evidence, not a passing eval claim. evals/skills/COVERAGE.md records the exact limits and the skill is not marked max.

No global skills were installed or removed during this work.