More and more people are using LLMs to write their documents: lawyers are using LLMs to write arguments full of bogus references; college students are submitting papers that make false claims.

LLMs are amazing, but they also hallucinate. People who skip the thinking get burned.

I don’t use the LLM to write my documents; I use it to critique them.

Generating a critique is as simple as pasting your text into the chat and asking for a critique. Example:

A critique of this: <paste text here>Code language: plaintext (plaintext)

Then I read the critique and rewrote; I repeat the loop if necessary.

Setting Up a Reusable Editor

Over time, my prompt got more complex, with more instructions. I wanted the spelling fixed, but the grammar and style critiqued without changing the text. I wanted my logical fallacies pointed out. The prompt was getting unmanageable.

I’m using Thaura.ai, which has a feature called Projects that lets you save prompts for a recurring task.

My first editor project used this prompt:

Extremely critical editor looks for all kinds of flaws, particularly flaws of logic, the regular list of logical fallacies, unusual or unconventional definitions, grammar errors, florid language, passive voice, failures to acknowledge limitations, ahistoricism, errors of fact, spelling errors, references to controversial sources, references to low quality sources, profanity, race stereotype, gender stereotyping, homophobia, insensitivity to issues of economic or social class, rudeness insolence.Code language: plaintext (plaintext)

That prompt caught plenty, but I wanted consistent scoring, so I added a system prompt to calculate a score for each flaw, plus an overall document score. The per-flaw scores told me what to fix first; the overall score gave me a number to compare against after each rewrite:

Reply listing each flaw, with a brief explanation, and a severity score of 1 to 3, higher is worse. Calculate an aggregate score of 0 to 100, using some kind of quadratic function, and advise on publishing the text.Code language: plaintext (plaintext)

The criticisms were often harsh, but they forced me to actually defend what I’d written, and the drafts got better for it.

Sometimes, the critique I got was brief, but even then, I forced myself to re-read the text and rework it. Doing a rewrite of a first draft usually improved it, but occasionally, reviewing my own work made me reconsider whether to publish it at all.

The whole loop is fast: paste, read the critique, revise. You don’t need to agree with every point — the value is in being forced to re-read your own words.

---

(This post was written with an editing assistant, which I’ll describe in a future post.)