Why is the second draft more important than the first?

The first draft is a thinking device. Its job is to find the argument, which usually surfaces late, once the writer has talked themselves into it. The second draft is the first pass that can see that argument whole.

How many drafts does a piece of writing need?

Fewer drafts than passes. A verification pass that traces every number to a source, a style pass, and a count taken from the rendered page each catch different faults, and none of them substitutes for another.

What is the difference between editing and rewriting?

Editing improves sentences inside an argument that is already fixed. Rewriting is what you do when the draft found a different argument than the one you set out to make, which is the normal outcome of a first draft.

Should you edit while you write?

No. Generating is cheap and the output sounds finished, while verification adds no new sentences and runs slowly, so in a shared sitting the generating half wins. We split them on a 255-page programmatic build for that reason.

How do you check a draft for invented facts?

Trace every number, date and capability claim to a named source file before publish, and delete anything untraceable rather than softening it. An invented figure reads exactly like a real one on the page.

Why does rewriting fail to fix some problems?

Because they were settled before the writing. A batch of twenty AI-written posts came back at around 595 words against a documented bar of 1,000 to 1,400, with no tables and no FAQs, because the standard was never in the brief.

The Second Draft Is the Real One

The first draft exists to find out what you think. The second is where the piece gets made. On a 255-page content build we ran generation and verification as separate passes, because when both happen in one sitting the confident half wins.

What is a first draft actually for?

To find out what you think. Most drafts get written before the argument exists, which is why the argument tends to show up late, after the writer has talked themselves into it. That paragraph is the real opening, and it is nowhere near the top.

A first draft is always bad as a finished piece. The only question worth asking of one is whether you can now say the claim in a sentence. If you cannot, another draft will not rescue it, because the problem sits further upstream than the prose.

The passes we run on a content batch, and what each one cannot catch:

Pass	What it produces	Mode	What it cannot catch
Plan	The claim and the structure	Decisive	Whether the claim is true
Generate	A complete draft, quickly	Confident	Its own invented detail
Verify	Every number traced to a source file	Adversarial	Rhythm and sentence shape
De-AI	Banned punctuation and AI sentence shapes removed	Mechanical	A thin argument
Measure	Word, table and link counts from the built page	Literal	Anything a count does not express

Why do generating and verifying have to be separate passes?

Because they disagree, and in a shared session the writer arbitrates their own case. On a 255-page programmatic build we split them deliberately: plan and generate, then verify and de-AI as passes of their own. The split is what made that volume survivable.

Generating is cheap and the output sounds finished. Verification adds no new sentences and runs slowly, so it loses whenever the two share a sitting. Given a pass of its own, it has no draft to defend.

What does an adversarial second pass check?

Every number, against a source file. Anything untraceable gets deleted rather than softened, and the check runs mechanically instead of as a careful read, because fluent output is confidently wrong and an invented figure looks exactly like a real one on the page.

Then the built page rather than the draft. Count the words, the tables and the internal links in the rendered output. A writer reporting that a piece meets the standard is reporting an intention, and intentions are what the first pass produces.

Which faults survive every draft?

The ones settled before any writing happened. A batch of twenty AI-written posts came back structurally clean and missed the documented content standard entirely: no first-party data, no tables, no FAQs, around 595 words against a bar of 1,000 to 1,400. Nothing in the brief had made the writer open the standard.

The second failure in that batch was quieter. After a repair pass, an adversarial audit still found invented supporting detail, because the source bank held one-line entries and the writers had closed the gaps themselves.

A second draft cannot add a table that was never planned. The faults that survive rewriting were settled in the brief.

When does a rewrite become a rule?

When the same fault turns up in consecutive drafts. Nine standing editorial rules were written for one newsletter at issue 20 rather than issue 1, because that is the point at which it became clear the same notes were being given twice. Eleven issues have shipped through that gate, numbers 20 to 30.

Forty-four content pages sit behind an automated validator that runs before publish. It arrived after the problem rather than before it, which is the usual order and the expensive one. A gate keeps catching a fault in drafts written by people who were never in the conversation.

The two checks that catch what a careful read misses are why a written style rule needs a script behind it and measuring the built page instead of the draft.