renwei-writing (人味儿写作)

An AI agent skill — “edit people’s words without erasing the person behind them.” Packaged as a portable skill (a SKILL.md methodology + supporting files, loaded via progressive disclosure), it targets a specific failure mode of generative editing: “AI 改完 之后…一遍比一遍缺人味儿” — each AI revision strips more of the author’s voice, converging on polished but impersonal prose. First-party project repo (tier T1). URL-only ingest.

What it is

• The skill (SKILL.md) — the methodology: definitions, editing rules, and documented edge cases for revising text while preserving the writer’s distinctive voice (“人味儿” = human essence).
• post-edit-checklist.md — a post-revision checklist adapted from Wikipedia’s “Signs of AI writing,” i.e. a verification pass that checks the output for tell-tale homogenization rather than trusting the edit.
• case-study.md — a before/after showing a failed (voice-erasing) revision and the corrected version.
• License: free for personal use; commercial closed-source use requires licensing.

Why it belongs here

• The augment pole, applied to writing. It’s a close cousin of lathe: the same primitive (loadable skills) aimed at the opposite of automation — not “AI writes for you” but “AI edits without overwriting the human.” Where Lathe keeps the human doing the work (hand-typed tutorials), renwei-writing keeps the human audible in the work. Both are the augment, not replace counter-current (augmenting-human-intellect, cross-wiki) — evidence that “skills” are pole-agnostic: the load-on-demand unit serves homogenizing automation and anti-homogenization alike.
• Output-grounded verification, again. Its post-edit checklist mirrors ai-job-search‘s “read the rendered artifact” discipline — verify the result (does it still sound human?) against a concrete rubric, not the model’s say-so. Here the rubric is Wikipedia’s AI-writing tells.
• A new skills host — Cola. It installs to ~/.cola/skills/renwei-writing — i.e. the Cola agent framework, a host not previously seen in this spoke (alongside Claude Code, ADK, Gemini CLI, Cursor, Codex, Kiro). Another SKILL.md-style directory convention is a fresh data point for the standardization-reach question: each new host either ratifies the portable-skill attractor or adds one more vendor-native ~/.<tool>/skills/ dir to the fragmentation.

Caveat

A single skill repo, not a benchmarked method — efficacy of the “preserve voice” rules is asserted via one case study. Cola itself isn’t sourced here yet (page it if more arrives).

The failure it names, demonstrated by a tool that promises the opposite (added 2026-08-03)

claudecode-writer is a Claude Code workspace whose central promise is that “your voice stays authentic across all platforms.” Its setup step is adding samples of your own writing, and its troubleshooting answer for “Claude doesn’t sound like me” is to add more. The intent is this page’s intent.

Its README’s own demo output is the counter-example: a hook (“still stuck in 2019”), a ❌/✅ ladder, one-line paragraphs, “Game changer.”, an engagement-bait closing question, a hashtag stack. Nothing in it identifies a person. That is 人味儿 removed at the generation step rather than at the editing step — the same erasure this skill was written against, reached from the other direction, and worth recording because voice-matching machinery did not prevent it.

Related

lathe · agentskills-spec · agent-skills · ai-job-search · claudecode-writer · augmenting-human-intellect · synthesis