Your first run
Ask for what you need: "write a LinkedIn post about...". Without a voice pack, Plume first asks for three things:
- three posts you've published and like, pasted or as a link it can open
- who you're writing for, in one line
- your main offer, with its link
That's about two minutes. Plume then writes your piece, ending with "Thin pack: 3 examples", and offers to keep what it learned as your voice pack, wherever you choose: files, a project, Notion, Google Drive.
For a complete pack
Ask for setup (/plume:setup in Claude Code). Bring 10 to 15 published pieces, comments included. Plume drafts your voice from them, then asks five questions at most, about what it can't see in your writing: your convictions, your readers, what's forbidden and what's only to avoid.
Give it something real
Plume won't invent your experience. A topic alone gets you up to three questions, not a draft.
Raw material is something you lived, saw or measured: a client story, a number, a sentence someone said. Notes in any order are fine.
Read the output
[figure: share of the team opening it weekly]: facts Plume didn't have. Fill them with real numbers, or cut the sentence. Never publish one as is.Ask "show the work"
To see what happened before you got the text:
- the lint report: AI tells, stock words, invisible characters, rhythm, formatting
- each critic's findings: the Reader, one of your readers who knows nothing about your rules; the Voice matcher, who only knows your published work; the Fact checker, who only knows your facts
- the editor's changes, and which version the judge kept
When a draft feels off, the critics' notes usually say why.
Post on LinkedIn, inform, EN, hook required, soft CTA: invite replies
{the piece, ready to paste, with [figure: ...] where a fact is missing}
To fill: the placeholders, listed
Thin pack: 3 examples. Share what you publish to add more.
Plume: lint, critiques and A/B choice available, ask "show the work".
The habit that makes it work
After you publish, show Plume the version you actually posted: "here's what I published". This is learn mode.
- Plume compares its draft with your final version and reports the edit distance: how much you changed.
- It proposes updates to your pack, each with the edits that justify it. You accept or reject them one by one. Nothing changes without your yes.
- It offers to add the published piece to your examples. In a thin pack, this is what makes the next draft closer.
Watch one number
If the edit distance doesn't go down over a few weeks, the pack needs work: run setup again with more pieces.
Pick the right mode
Your voice pack
- Where: anywhere your assistant can read: attached files, a project, a
plume-pack/folder, Notion, Google Drive. Any readable format works; Markdown is the easiest to keep up to date. - What matters most: your published pieces. Then your convictions (what you believe, what you push back on) and your bans.
- List markers and symbols for plain-text platforms like LinkedIn: set them once in
formats.md, underglyphs.
Keep it private
Your pack holds your positioning and your numbers. Share the skill, not the pack.
When it goes wrong
It asks questions instead of writing
Your request had no raw material. Add one concrete thing: a story, a number, a sentence someone said.
It doesn't sound like me
Check the "Thin pack" line and add published pieces. Then use learn mode after each post.
The lint flags a word I really use
Add it to stock_words_allow in your pack's formats.md.
The result says "Reduced Plume"
Your platform can't run independent critics, so they ran one after another. Use Claude Code for the full pipeline.
It says "lint done by the model"
Code execution is off, so the checks were done by reading. In the Claude app, turn code execution on.
Will it post for me?
No. Plume writes; you publish. It never posts or sends anything for you.
Ready for your first draft?
Deciding, not writing? Prism runs your decision through a council.
