[christso]

Summary

• Add the productivity plugin and register it in both native catalogues.
• Keep write-naturally general, with conditional references for reviews and technical designs.
• Add a short README about long-conversation drift, fresh skill reads and carrying clarifications into final text.
• Add 28 fictional public cases, including eight inspired by Humanizer. No captured private writing or review provenance is included.
• Add a comparison without the skill, using the same model and graders. The baseline rejects recorded tool calls and missing traces; it cannot rule out instructions supplied before the model turn.
• Upgrade to Promptfoo 0.124.1 and explicitly pin Codex SDK 0.162.1. The writer defaults to gpt-6.1-sol; the grader remains gpt-5.6-luna.

Checks

Catalogue and Markdown-link validation, skill validation, config validation, privacy inspection and whitespace checks pass. Bun tests: 21 pass. Independent review found no remaining implementation or privacy issues.

Behavioral results

Run	Pass	Fail	Errors
Full public suite	21	7	0
Humanizer with the skill, focused run	8	0	0
Humanizer without the skill, first run	8	0	0
Humanizer without the skill, repeat with full trace checks	8	0	0

Style calibration accepts eight good controls and rejects eight bad controls. It initially missed repeated context in H6; the rubric now explicitly honors what the recipient already knows. Formatting requirements and factual requirements are separate for the Humanizer cases.

Two failures in the full run were grader errors: C2 and H4 were judged using wording present only in the source draft. Added an explicit candidate/source boundary and calibrated it with four controls: both valid outputs pass, and both faulty controls fail. The full suite has not been rerun after that grader correction. The Humanizer repeat therefore has seven raw passes and one verified false rejection, rather than a writing failure.

Five genuine full-run failures remain: C4 repeats discarded review context; D2 and D3 omit supplied constraints; D7 drops a required identifier; C11 omits the original workshop time. Some inherited readability scores also repeat factual failures, so per-metric scores still need care. C7 and C12 pass with the upgraded writer, including a separate two-case smoke run; earlier luna runs had failed those cases.

These eight short rewriting cases provide no evidence that Humanizer-specific skill instructions are needed. The comparison does not establish behavior for every pattern, writing from scratch or long conversations. No Humanizer-specific guidance was added. Model outputs stay local and are not committed or shared.