Free skill for Claude and ChatGPT

Plume

Writes like you. Reads like your audience.

The writing desk for people who publish under their own name.

In short

  1. Plume writes, edits and repurposes LinkedIn posts, articles, emails, messages and bios, starting from pieces you actually published.
  2. Three critics check every draft, and each one misses something on purpose: a reader from your audience, a voice matcher, a fact checker.
  3. It invents nothing. A missing figure becomes a placeholder for you to fill.
  4. Free and open source. Full Plume in Claude Code, reduced mode in ChatGPT.

Most AI drafts sound like the same person. Plume starts from what you already published.

For

  • Freelancers and consultants who post under their own name
  • Coaches, trainers, founders, and anyone rewriting their LinkedIn profile
  • People with a few published pieces they are proud of
  • One voice, or one brand with its own pack

Not for

  • Writing from nothing ("give me 10 post ideas")
  • Engagement bait and growth hacks
  • Translating, or writing as someone without their pack
  • Designing visuals

How it works

Your material goes in, a piece you'd sign comes out.

01 / 07

Bring raw material.

Something you lived, saw or measured: a client story, a number, a note. Without it, Plume asks up to three questions. It never fills the gap itself.

02 / 07

Plume sets the brief.

Format, platform, intent and language, inferred and stated in one line. Change any of them before it writes.

  • FormatsPost, article, email, message, comment, slides, script, page, bio.
  • PlatformsLinkedIn, Reddit, Slack, Teams, WhatsApp, blog, email.
03 / 07

Three angles, you pick one.

Three finders work in parallel, each with one lens, none sees the others.

  • What it costsThe price of the problem, in time, money or nerves.
  • What it contradictsThe common belief your material pushes against.
  • The lived momentThe scene, the day, the sentence someone said.
04 / 07

The writer starts from your examples.

Three to five pieces you published, in the same format and language, come first. Your rules come second. Then a script checks what a script can check.

  • LintAI tells, stock words, invisible characters, too few concrete details, rhythm, formatting the platform does not render. It flags, it never scores.
05 / 07

Three critics, three blind spots.

Each one lacks something the writer has, so it cannot share the writer's blind spots.

  • ReaderOne person from your audience, scrolling. Never sees your rules or your intent.
  • Voice matcherOnly sees what you published. Flags what doesn't sound like you.
  • Fact checkerOnly sees your facts. Flags anything it can't trace.
06 / 07

One edit, one blind pick.

The editor fixes what the critics flagged and nothing else. The judge compares both versions without knowing which one is the revision, and keeps the better one.

  • JudgeSometimes it keeps the first version. That is the point.
07 / 07

You publish, Plume learns.

Show Plume what you actually posted. It compares, proposes updates to your pack with the evidence, and writes only what you approve.

Who sees what

Each critic is blind to something, on purpose.

A critic that knows everything the writer knows shares the writer's blind spots. So the Reader never sees your rules, the Voice matcher never sees your rules either, and the Fact checker only sees your facts.

What each role can see
Your rules and bansYour published piecesYour factsYour audience
WriterSeesSeesSeesSees
ReaderNever seesNever seesNever seesSees one persona
Voice matcherNever seesSeesNever seesNever sees
Fact checkerNever seesNever seesSeesNever sees

Everyone also sees the draft. The fact checker also sees the raw material you gave.

A marked draft

From your notes to a post you'd sign.

Sample, to replace with a real run

A client asked me to rebuild their CRM. For the third time.

The first two versions worked. Nobody opened them.

In today's fast-paced world, a tool is only as good as the habits around it.

So this time I started with one question: what does Sarah check every Monday at 9?So this time I started with one question: what does their sales lead check every Monday at 9?

We built that view first. Three properties, one filter. Nothing else.

It's not about the tool, it's about the people.

Six weeks later, figure: share of the team opening it weekly of the team opened it every week.

What do you think?Start with someone's Monday morning. The tool comes after.

Plume's first draft, from your notes and five of your published posts.

Brief

LinkedIn post, intent inform, angle: the lived moment. One CTA at most.

The lint runs first. Code, not opinion.

Line 3

Generic opener.

Line 6

"Not X, it's Y."

Line 8

Reflex question as a closer.

Three critics, each with only part of the picture.

Reader

Line 1 stops me. Line 4 loses me: who is Sarah?

Voice matcher

Line 3 isn't you. In your posts, the lesson comes from the scene, never from a maxim.

Fact checker

"Third time" and "three properties" are in your notes. The usage figure isn't: kept as a placeholder.

One round of minimal changes.

Editor

Four changes, each one flagged by the lint or a critic. Nothing else touched.

The judge picks without knowing which is which.

Judge

Kept B, the revision. Shorter, and the lesson comes from the scene.

Before you publish

One placeholder to fill with your real number.

Your voice pack

Plume is only as good as what you give it.

Your voice, your bans, your readers, your facts and, above all, pieces you actually published. Setup builds the pack with you, or imports the voice guide you already have. It lives in your files, outside the skill, so an update never overwrites it.

Keep it wherever you like: a Notion page, a Google Doc, a PDF, plain files. Plume reads your pack.

Keep it private: it holds your positioning and your numbers.

Get the Notion voice pack template

Free. Duplicate it, fill it, point Plume to it.

voice.md
Who you are, what you believe, how you sound.
anti-patterns.md
What never appears, with weak and strong examples.
lexicon.md
Your expressions and signature lines.
audiences.md
The specific readers the Reader plays.
facts.md
Offers, prices, links, proof. The only source for claims.
formats.md
Your overrides: shorter posts, extra bans, signatures, your list marker and symbols.
examples/
Pieces you published. The part that matters most.
learn-log.md
The updates you accepted, with their evidence.

Measured

One number tells you if it works.

Edit distance to publish: how much you change Plume's draft before you post it.

After each piece, show Plume what you actually published. Learn mode compares the two, reports the number, and proposes changes to your pack with the evidence. You approve them one by one. If the number doesn't go down over a few weeks, the pack needs work, and you'll see it.

Install

Pick your platform.

  1. Screenshot: GitHub release, ZIP
    Download the ZIP.
  2. Screenshot: Customize, Skills
    Open Customize > Skills.
  3. Screenshot: Upload
    Upload the ZIP.
Download the ZIP

Turn on code execution, so the checks can run. Then say "set up my Plume pack".

Run these two commands in Claude Code, then start with /plume:setup.

/plugin marketplace add The-Discipline/skill-plume
/plugin install plume@plume

Each mode has its command: /plume:write, /plume:edit, /plume:repurpose, /plume:review, /plume:learn. Or just /plume, and Plume picks the mode.

Beta

ChatGPT desktop app

Open Plugins > Skills, upload the Plume ZIP, then call @plume. On the web and mobile apps, ChatGPT only runs skills packaged as plugins: not available yet.

Codex

Copy the plume folder to ~/.agents/skills, then call $plume.

Reduced mode

ChatGPT doesn't run independent agents for skills. The critics run one after another in the same conversation, each restarting from only what it may see, and the result is labelled "Reduced Plume". For independent critics, use Claude Code.

Your voice pack is plain Markdown. Paste voice.md, anti-patterns.md and a few published pieces into a Gemini Gem, a project's instructions or the start of a conversation.

What you lose

The critics, the checks and Learn mode. You get a writer that knows your voice, not Plume.

Pack templates on GitHub

Start with three posts you're proud of

Setup builds a first pack from three published pieces. It's a thin pack, and Plume says so. Every piece you publish after that can join it, and the drafts get closer.

New to Plume? Read Get started for your first run.

FAQ

Questions before you install.

Is it free?

Yes. Plume is free and open source.

Do I need to prepare anything?

No. Without a pack, Plume asks for three posts you've published, who you write for and your offer, then writes your piece. It offers to keep what it learned after.

Will it sound like me on day one?

As much as your pack allows. Three published pieces are enough to start, and the output tells you the pack is thin. With 10 to 15, it starts close. Learn mode brings it closer with every piece you publish.

I haven't published anything yet. Can I still use it?

Yes, in a neutral voice, clearly labelled. Come back after five posts: that's when it starts to sound like you.

Does it post for me?

No. It writes, you publish. It never connects to your accounts: most platforms only allow automated posting through approved apps, and it is your account that takes the risk.

Will it invent figures or client stories?

No. Claims come only from your raw material and your facts file. Anything missing becomes a placeholder for you to fill.

Does it write in French?

Yes. It writes in the language of your piece, from your published examples in that language.

Can it write for my company?

Yes, with a voice pack built for that brand. It is made for one voice at a time.

Does it work in ChatGPT?

In the ChatGPT desktop app and in Codex, in reduced mode: the critics run one after another instead of independently, and the output says so. Claude Code runs the full Plume.

Where does my voice pack live?

In your own files or your Notion, outside the skill. Plume reads it, and only writes to it when you approve a change.

Credits

Built to replace two of my own projects.

Origin
I wrote with one AI project and reviewed with another. They shared the same context and drifted apart. Plume puts both in one place, with critics that can't drift because they can't see.
Sister skill
Prism, the decision council for businesses of one.
Built by
Virgile Gouala. I build Notion systems for freelancers and small teams, at goua.la.