sharp-pen review

---

Focused proofreading, run by your AI assistant. It fixes your writing without changing your tone or reformulating what you wanted to say. Hand it a draft and it hands back a review page: your text on the left, suggestions on the right, every difference clickable. Accept a change with one click; when there is more than one reasonable phrasing, pick from a dropdown. Nothing is applied until you say so. It works in any language the model you choose supports, and detects the language from the text on its own.

https://github.com/user-attachments/assets/7fab5412-2428-4cd1-8e86-4208745bcfe7

It is packaged as an agent skill and plugin: a folder with instructions and two small Python scripts that work in Claude, ChatGPT, Codex, Copilot CLI and other clients that read the skill format.

sharp-pen is also available on the VS Code Marketplace for reviewing Markdown and plain-text editors with a user-installed local AI CLI.

Two passes, toggled in the page:

• Level 1 — spelling and grammar. Things that are wrong: typos, agreement, prepositions, missing articles, homophones, comma splices.
• Level 2 — sentence construction. Fragments, broken parallelism, buried points, sentences that trail off. One entry per sentence.

The left pane is always your text byte for byte. The skill suggests form, not substance: it never changes what a sentence claims, strips a hedge, or makes casual prose corporate.

Why

Many writers want to do the rough draft themselves. Getting the words down is part of the thinking, and handing it to a model takes that away. What they want afterwards is focused copy editing: catch the typos, fix the grammar, flag the sentence that doesn't quite work, and leave everything else alone. Asking a general assistant for that is surprisingly hard. Tell it to proofread and it rewrites, smooths out the voice, drops a hedge the author meant, or quietly changes a claim. And it hands back one version, take it or leave it. sharp-pen constrains the model to form, not substance, and splits its suggestions into hard errors and optional sentence fixes. Where more than one phrasing would do, it offers two or three alternatives rather than picking for you, so the choice of wording stays with the author. Every suggestion is a separate click. The author stays the author. The AI is a careful copy editor that never touches the text without asking.

Languages

sharp-pen works in every language that your AI provider of choice, and the model you pick there, supports. There is nothing to set: it works out the language from the text itself, reviews by that language's own rules, keeps its spelling variant (US or UK English, Brazilian or European Portuguese), writes its notes in it, and never translates. A draft that mixes languages is reviewed passage by passage. How good the review is in a given language depends on the model; widely used languages get the most reliable results.

Try it

No install needed: the live demo is a 400-word blog draft reviewed by sharp-pen. Toggle between Level 1 and Level 2, click a highlight to accept it, open a dropdown where more than one phrasing fits, then download the result.

VS Code extension

Install sharp-pen from the VS Code Marketplace to review Markdown and plain-text files without leaving VS Code. The extension keeps the source editor open beside an interactive preview, where you can inspect suggestions, choose alternatives, and apply only the changes you want.

It runs your chosen local AI CLI on the desktop or remote workspace extension host. See the extension README for supported clients, model selection, remote setup, and privacy details.

Install

You don't need a terminal. sharp-pen is free on every Claude plan, including Free, and on ChatGPT desktop.

Claude — web or desktop app

1. Download sharp-pen-latest.zip.
2. In Claude, open Settings → Customize → Skills.
3. Click +, choose Create skill, then Upload a skill, and pick the zip.
4. It appears in your skills list, switched on. Paste a draft into any chat and ask for proofreading.

Works the same in Cowork. Older versions are on the Releases page.

ChatGPT desktop / Codex

codex plugin marketplace add Nobbettt/sharp-pen
codex plugin add sharp-pen@sharp-pen

Restart ChatGPT desktop after adding the marketplace. Then type $sharp-pen followed by your text, or just ask for proofreading.

For a standalone Codex skill instead, unzip sharp-pen-latest.zip into ~/.agents/skills/, or .agents/skills/ at a project root.

Claude Code

/plugin marketplace add Nobbettt/sharp-pen
/plugin install sharp-pen@sharp-pen

Then paste a draft into any project and ask for proofreading, or run /sharp-pen:review followed by the text.

Other clients (Copilot CLI, …) — unzip the same release into your client's skills directory so it contains a sharp-pen/ folder with SKILL.md at its root.

Requirements: Python 3. The scripts use only the standard library.

Use

Paste or point at a draft and ask for proofreading, or type sharp-pen followed by the text:

/sharp-pen please look over this before I publish it:

Last month I set myself a challenge: build a small side project in one
weekend. Have to say it took alot longer then that.

The assistant saves your text untouched, works out the suggestions, builds the review page and opens it for you. Click through the changes, pick the phrasing you want where there are alternatives, and download the finished text from the page. Your original is never edited in place.

In the ChatGPT desktop app the review opens in a browser pane beside the conversation:

In the Claude desktop app the review page opens right beside the conversation:

Privacy

The sharp-pen service collects, stores, and transmits nothing. The review page is a single self-contained HTML file with no external resources; the local server binds to 127.0.0.1 only. Your draft is read by whichever assistant you run the skill in, and a selected CLI sends review prose to its provider under that provider's terms. sharp-pen adds no transmission of its own. Details in PRIVACY.md.

About the author

I'm Norbert Laszlo, an AI solutions architect with a background in data science and software engineering. I build tools for working with AI agents and write about evaluation, agents and practical AI product work at norbertlaszlo.com and on Medium.