editor

Files (27)

• hero.gif · .github/assets/
• review-table.png · .github/assets/
• .gitignore
• AGENTS.md
• CLAUDE.md
• LICENSE
• Makefile
• README.md
• SKILL.md
• dispatch.sh
• customisation.md · docs/
• troubleshooting.md · docs/
• package.json
• code-style.md · references/
• cute-labels.md · references/
• effective-shell-style.md · references/
• hacker-news-effective-shell.md · references/
• linkedin-effective-shell-posts.md · references/
• linkedin-post-examples.md · references/
• obnoxious-humility.md · references/
• output-style.md · references/
• review-examples.md · references/
• substack-signal-noise-example.md · references/
• writing-style.md · references/
• hero-demo.txt · scripts/
• hero-play.sh · scripts/
• hero.tape · scripts/

SKILL.md (first lines)

# Editor

Edit or review a file to remove overly stylised AI idioms and other patterns that add noise or cognitive overhead. By default, the skill runs the review through GPT via Codex, which can expose structural patterns that the current session's model misses.

## Examples

```
/editor "clean this up"
/editor gpt "what are the top 10 things you'd rewrite in qb-scholar. dont change"
/editor gpt "how would you rewrite the skill md"
/editor "don't rewrite, tell me what is wrong"
/editor claude "simplify section 6"
/editor gpt "restructure this file into Forbidden/Discouraged/Preferred. writing-style.md"
/editor quick
```

The target file is usually clear from conversation context or named in the instructions. The model defaults to `gpt`.

## Arguments

`/editor [model] "<instructions>"`

- **model** - selects a model. Defaults to `gpt`. Available: `gpt`, `luna`, `o3`, `claude`. If the first word does not match a model alias, treat it as the start of the instructions and use `gpt`.
- **quick** - clean the file in the current session, described under "quick mode" below. Takes no model.
- **instructions** - what to do. If the instructions contain "list", "check", "review", or "don't rewrite", return issues only. Otherwise rewrite the file.

## Write the response in the style too

The rules apply to your own prose, not only to the file you are working on. Reviews, summaries and rewrites all use characters available on a standard keyboard: `-` for dashes, `"` and `'` for quotes, `->` for arrows, `...` for ellipsis. A review that flags an em dash while using one has failed its own check.

## Say who did the work

End every response, in every mode, with a single line naming the harness, how it ran, and the exact model. Take the model from the flag `dispatch.sh` passed to the CLI. Never guess it.

```
Editor: Codex, dispatched as a new process, model gpt-5.6-sol
Editor: Claude, current session, model claude-opus-5
```

Name the harness you are actually running in. `quick` always runs in the current session, so it reports the current harness and model.

## quick mode

`/editor quick` is for a fast pass when the copy does not have to be right, only free of the worst noise. Do it in the current session with whatever model and harness is running, and do not dispatch.

- Read the Forbidden section of `references/writing-style.md`. Skip the other references.
- Fix the violations in the file, plus any obvious filler.
- Leave anything that needs judgement, and say so in one line.

Write the file and give a one-line summary of what was cut, then the harness line. For a careful pass, use the default `/editor` instead.

## What to do

These steps cover review and edit mode. For `quick`, use the section above.

1. **Identify** the target file from the instructions or conversation context.

2. **Write a prompt file.** Write it to the scratchpad directory. Include:
   - The content of `writing-style.md`, `review-examples.md`, `cute-labels.md` and `obnoxious-humility.md` from this skill's `references/` directory.
   - The content of `code-style.md` when the target is code (source files, or prose that documents code such as comments in a diff).
   - The content of `effective-shell-style.md` only when the instructions mention that tone.
   - The content of the target file. For code targets, include the code itself, never extracted comment lines alone: judging a comment needs the code it annotates in view.
   - The user's instructions.
   - This postscript:

     For edit mode:
     > Apply the Forbidden rules: fix every violation. Apply Discouraged rules
     > where the pattern is clearly noise. Apply Preferred rules where you have
     > high confidence the change improves the text.

     For review mode:
     > Flag every Forbidden violation with the line, the problem, and a fix.
     > Flag Discouraged patterns where they add noise. Suggest Preferred
     > improvements where they would strengthen the text.

     For both modes, append:
     > Write your reply using only characters on a standard keyboard. No em
     > dashes, en dashes, curly quotes, arrows or ellipsis characters, in your
     > own prose as well as in any text you rewrite.

     For code targets in edit mode, append:
     > Apply the code style rules: fix every violation.

     For code targets in review mode, append:
     > Flag every code style violation with the line, the problem, and a fix.

3. **Dispatch.**
   - If model is `claude`, run directly in the current session. No dispatch script needed.
   - Otherwise, run `dispatch.sh` from this skill's directory:
     ```
     dispatch.sh <model> <prompt-file>
     ```
     The dispatch script handles harness-specific CLI details. To add a new model or harness, edit `dispatch.sh`.

4. **Present results.**
   - **Review mode:** show a numbered issue list with line/quote, problem, and fix.
   - **Edit mode:** show a brief summary using relevant before/after examples. If the user approves the changes, write the file.

End with the harness line, as described above.

From dwmkerr/editor, MIT.

More from dwmkerr/editor