Restore Unslop as a focused prose editor

[ariobarin]

The existing skill mixes useful AI-writing diagnostics with instructions to invent personality and blanket style bans. This consolidates the pattern catalog for general prose editing, treats patterns as contextual clues, and keeps website-specific voice preferences out of the general guidance.

[ariobarin]

Unslop editing check

September 5, 2026. This supersedes the earlier public-account-only revision and its checks.

PR #328 retains Unslop's purpose as an editor for recognizable AI-writing problems in explanations, documentation, PR text, and public prose. Seven categories consolidate the original catalog: empty praise, vague authority, inflated language, mechanical structure, redundant presentation, false variety, and canned rapport. They are contextual clues, not automatic bans. Useful lists, contrasts, technical terms, and working sentences should survive. The skill no longer asks the agent to invent personality. Detailed website voice preferences remain in the website's editorial guidance. The PR adds no global instructions.

Comparison

Two fresh Codex CLI runs used GPT-6 Astra at medium effort, the same working directory, three synthetic editing requests, no tools, and ignored user configuration. The baseline received no candidate instructions; the candidate received the revised skill. This was a prompt-level comparison within Codex's default context, not a bare-model experiment or an automatic skill-loading test.

Both versions:

• Reduced a promotional release note with invented authority and unsupported performance claims to the two established facts: uploads resume after connection interruptions, and the public API is unchanged.
• Retained the justified three-step numbered recovery procedure in the correct order, including the seven-day retention detail.
• Preserved "API surface", the compiler/schema mechanism, the distinction between JSON and CSV, and the limits on the reported 12% improvement: three sample files and no large-input testing.

There was no substantive advantage over the default on these cases. This is a limited check against destructive editing, not proof that the full catalog improves quality or that any untested category is unnecessary.

Description check

A separate text-only run classified eight requests from the description. It selected the inflated release-note edit and an explicit request to unslop a formulaic article. It excluded a voice-preservation-only website edit, a code fix, verbatim quotation, a new technical explanation without an identified writing problem, a file rename, and implementing validation from a voice note. The website's editorial instructions own its voice-preservation task. This checks the description's stated scope, not end-to-end discovery in every client.

The skill validator, whitespace check, PowerShell 7 portable suite, and Windows PowerShell 5.1 portable suite passed. No recurring tests were added for this editorial revision.