The mistake I keep watching people make is simple: they open ChatGPT — or whatever the current chat box is — and ask it to write the article for them.
That is not a drafting shortcut. That is handing your byline to a model.
Writers asking “how do I start?” often don’t need another prompt. They need an editorial process. Then the prompt pack arrives, and the process never does.
That distinction is why I built Pysar.
The fluent mush problem
Open a feed. Smooth sentences. Polite transitions. Nothing wrong enough to call out — and nothing yours enough to care about.
You know the feeling before you finish the first paragraph. Readers bounce. Trust thins out. Plagiarism panic is the wrong headline for that.
Better prompts do not fix a missing process. They just make the emptiness arrive faster.
The trap: “just write it for me”
Someone wants an article on why remote work failed their team. They paste the topic into a chat and ask for a finished piece.
Fluent comes back. It might even score well on shallow quality checks. It will not carry their fights, their evidence, or a voice anyone who knows them would recognize. Ask where a claim came from — often nothing to point to.
A make-this-better button that invents confidence without sources is the opposite of editorial work.
Why build another AI writing tool?
Because most of them share the same abstraction: a chat UI that collapses the whole editorial process into one prompt → one output.
State disappears into the scrollback. Voice is a vibe in the system prompt. Sources are whatever the model asserted. Revisions overwrite each other. Publish is “copy the answer.”
I spent years in newspapers, a publishing house, and tech-book reviewing. One stubborn rule survived: the author stays accountable for the take. An editor can challenge structure, demand sources, and push for a sharper claim — without becoming the author.
So I built software around a different abstraction. Writing as a stateful pipeline, not a single completion:
brief → research → sources → draft → revision → review → publish
Each stage leaves artifacts you can open, argue with, and veto. The model does the heavy lifting. The take stays yours.
Treat AI as a writing collaborator, not as the author. If you still chose the thesis, kept the voice, and can defend the sources, the piece is yours — even when AI helped. Outsource the take and it isn’t, no matter how pretty the paragraphs are.
No — you are not the unpaid intern
Someone recently asked why they should have to do the work while AI “simply directed traffic.” Fair ask. That is the point.
“Collaborator, not author” gets misheard as you still do the grind, AI gently steers. Backwards.
Autopilot runs the editorial chain end to end — you are not rewriting every pass by hand.
What stays yours is not the keystrokes. It is the take: the claim you will defend, the voice, the sources, the veto, the decision to publish. Direct traffic at the decision layer. Let the model do the grind. Keep those decisions visible — not buried under one magic button.
What that looks like
Pysar is an open-source, author-directed editorial engine for that rule. Idea or rough draft in; a complete draft out. It does not post for you. It runs as a CLI, exposes an MCP interface, and ships with ps-* skills for Claude Code, Cursor, and Codex.
Voice, sources, instructions, and revisions stay explicit.
Same remote-work idea, different path.
Instead of:
“Write an article about why remote work failed my team.”
You start with:
“Remote work did not fail because people stopped collaborating. It failed because our team lost the informal coordination that used to happen around the office.”
Then the system can research the claim, challenge it, find supporting sources, draft against the brief, and revise — without silently changing what you actually wanted to say.
That is the difference: AI does the grind. You own the take.
More explicit structure from AI can make the writing more human, not less. Hiding decisions is what produces sameness.
Docs, essays, research write-ups, articles, management notes — the same approach works anywhere you have a take. Already got a messy draft? Start from that text. Don’t throw it away for a fresh chat rewrite that sounds like everyone else.
Docs: getpysar.com/docs
Longer essay on Medium: AI should help you write — not replace your voice
Issues and ideas welcome on github.com/malyshev/pysar.
Do not rent a voice to say what only you can say.
Top comments (0)