Start with the author’s draft and give a reason for every change

Renwei Writing

Edit Without Losing the Voice

Lightly edit personal articles, first-person accounts and transcribed speech. Find genuine reading obstacles, preserve perspective and cadence, and compare an over-polished version with the original to recover useful choices.

By 橘子 (Orange)

Renwei Writing: Edit Without Losing the Voice project image
GitHub stars
1,111
Checked

Start with a short passage and judge each edit

Read the original and revision aloud. Check whether obstacles are removed, hesitation and humor remain, and every change has a reason. An unchanged draft or a reverted edit can be the right outcome.

Pick a task and get started

Have the assistant read the principles and checklist before choosing a task. The first task contains a short demonstration. For your own work, replace it with a draft and purpose, keeping versions that can be compared.

Start from a draft, a polished version or a set of edits

Use a supplied demonstration to preserve a conversational account while clearing real obstacles.

  1. Keep the complete material and confirm access

    The author clones the complete repository into ~/.cola/skills/renwei-writing for Cola. Use your own host’s documented skill directory for automatic discovery, preserving the complete folder. Alternatively, download the material and explicitly ask the assistant to read SKILL.md and its checklist in the current session; that does not verify automatic loading.

    Read the author’s setup guide

    First use: download, then read in order

    The commands below download the repository into a local renwei-writing folder and select the commit reviewed for this page. It does not install the skill in every host; do not overwrite an existing folder. Ask the assistant to read renwei-writing/SKILL.md and both files under renwei-writing/references, then confirm the order: decide whether to edit first, check changed sentences afterward. The author documents a Cola directory; use your host’s actual convention for automatic discovery.

    git clone https://github.com/orange2ai/renwei-writing.git renwei-writing
    git -C renwei-writing checkout c6566884997f3688478f43ecdac29423757548a0
    Read how the principles and checklist work together
  2. Give your AI this first task

    These editorial prompts follow the author’s method. Demonstration text and filenames are suggested conventions, not results obtained by running a model. Compare actual revisions with the original.

    Use renwei-writing. Read the local renwei-writing/SKILL.md and renwei-writing/references/post-edit-checklist.md first. If they are not downloaded, start with the pinned author document at https://github.com/orange2ai/renwei-writing/blob/c6566884997f3688478f43ecdac29423757548a0/SKILL.md and follow its checklist link. Do not infer the method from its name. This is editorial demonstration text, not my personal experience: “Yesterday I checked the signup page, I'd found the problem but I kept going through the logs, afraid I'd missed something again. Then a colleague pointed to one line and said, let's fix this first. That's when I stopped.” The audience is familiar colleagues and the purpose is a short reflection on work. Keep the first-person perspective, hesitation and final pause. Add no technical explanation or identity for the colleague. Identify genuine reading obstacles before editing. Make only changes you can justify, without a quota or percentage. Return the draft unchanged if that is sufficient. The underlying checklist is adapted for Chinese; judge English syntax and cadence in context. Return the original, the revision and an original→revision→reason list. Apply the post-edit checklist only to changed sentences. Check that no scene, number, experience or grand conclusion was added. Mark uncertain changes as reversible and explain what to listen for when reading aloud.

  3. Check what you receive

    • The original, a complete lightly edited passage and a reason for each change.
    • A comparison of speaker, facts, emotional nuance and cadence; greater polish is not automatically an improvement.
    Read the author’s full guide
More about this skill

Orange and Cola developed Renwei Writing after repeated polishing made a draft feel less like its author. The skill addresses editing existing writing: a personal article, an account of an experience or transcribed speech already supplies the material. The assistant helps readers follow it while preserving the writer’s choices.

The author describes perspective, lived stakes and individual traces as useful editorial cues. They include where a speaker observes from, the experiences behind a judgment, and particles, pauses, repetition and uneven sentence lengths. These cues need context. Adding a place, time or emotion does not manufacture lived experience, and an original draft can still contain factual errors.

The work has two layers. Read SKILL.md to decide whether a change is warranted, then use the post-edit checklist on sentences changed in this round. It examines inflated significance, promotional language, formulaic syntax, formatting, tone and filler. The scope matters: this is not an instruction to standardize the author’s entire text.

The published case places an original, a failed rewrite and an accepted revision side by side. The failed version introduced unsupported imagery and numbers and removed vocal nuance. The accepted version mainly adjusted local wording and punctuation. Its three edits describe that case; they are not a quota for every future article.

The deliverable should make every choice inspectable: what changed, what obstacle it addressed, why this wording helps, and what can be restored. The repository supplies an editing process, not a fact-checking engine, authorship test or humanity score. The writer still decides whether the revision says what they mean.

Sources and verification

Sources checked

Read all six files at the pinned commit: both READMEs, the skill, post-edit checklist, case study and license. No skill installation or model run was performed. Editorial examples explain the process and are not tested editing outcomes.