Keep One Draft From Before the Suggestions

When I first tried to improve Aethel after a low-value-content rejection, the obvious response was to make every article look more complete.

The essays became longer. They gained sources, practical sections, limitations, and polished conclusions. Each individual page looked defensible. Read in sequence, they sounded as if one invisible template had been asked to discuss twelve subjects.

The failure was not grammar. It was provenance. I could no longer see which sentence had begun as an experience, which had begun as a research claim, and which existed because fluent prose seemed to require another balanced paragraph.

I needed one artifact from before the suggestions—not because solitude makes writing pure, but because comparison requires a baseline.

A pre-assistance draft is evidence, not a sacred object

The draft can be rough. It may contain a scene, a question, a diagram, or a list of claims. Its job is to show what the writer noticed before alternatives crowded the page.

Without that artifact, the first generated structure can become the default. Later revisions may be thoughtful, but the writer has no record of whether the tool expanded the inquiry or replaced its center.

This is especially important for work that carries a name. Authorship is not a keyboard census. It is responsibility for selection, evidence, interpretation, and consequence.

The U.S. Copyright Office’s report on copyrightability addresses human authorship in a legal context. Google’s guidance on generative AI content focuses on quality and usefulness rather than banning a production method. Neither source establishes my proposed workflow. They clarify two boundaries: human contribution matters, and generated text is not automatically low quality. The remaining editorial question is how a writer can make their own contribution inspectable.

The three-artifact archive

For a substantial essay, keep three compact records.

Artifact	Contents	Editorial use
Origin note	The initial scene, question, claim, or diagram made before generation	Shows what prompted the work in the author’s own terms.
Assistance ledger	Important suggestions accepted, rejected, or verified	Shows how tools changed the piece.
Responsibility draft	The final argument with a source-to-claim map and unresolved limits	Shows what the author is prepared to defend.

The archive can remain private. Readers rarely need a full production log. The point is to preserve enough evidence for the author to review the process honestly.

Example from an Aethel essay

Origin note: “Adding more memory made a simple chat response slower. Why did a convenience feature begin to control which past facts entered every answer?”

Assistance ledger: A tool helped compare privacy frameworks, identify overlapping sections, and test alternative structures. A suggestion to make a broad claim about AI memory damaging identity was rejected because the cited research did not establish it.

Responsibility draft: The final article argues that memory should be governed through scope, provenance, expiry, influence, correction, and derivative deletion. The product-building experience supports the design problem; external sources support privacy and governance claims; the identity interpretation remains explicitly editorial.

That record is more useful than declaring the article “human-written” or “AI-assisted.” It shows where judgment occurred.

Specific failure is harder to mass-produce

The strongest material in Aethel came from decisions that did not make me look uniformly competent: building an agent path so elaborate that a simple greeting became slow; allowing memory retrieval to increase time before the first response; adding broad functionality before the product’s explanation was clear; discovering that a technically polished publication still sounded templated.

Generic prose tends to remove those irregularities. It turns failure into a smooth lesson and gives every section the same emotional distance.

A personal detail is not valuable merely because it is personal. It must change the argument. The slow greeting matters because it exposed a mismatch between architectural ambition and user intent. The templated library matters because it showed that formal completeness can coexist with weak editorial identity.

Before publication, each experience should be checked by the author. The codebase proves that the site and seed contain these claims. It cannot prove that the events happened as described.

Solitude is not accessible or universally productive

A requirement to write alone can become exclusionary. Some people think through dialogue. Translation, dictation, autocomplete, and assistive technologies may make composition possible. A beginner may need an example before they can form a useful question.

The protected artifact therefore need not be unaided prose. It can be the person’s selected evidence, spoken account, annotated example, or decision about what the work must accomplish.

The boundary is agency: before the system supplies a complete answer, is there a trace of what the person noticed, valued, or intended to test?

For some tasks, no protected zone is needed. Routine summaries, administrative messages, and transformation work may reasonably begin with automation. The practice is most valuable when the work claims original judgment or when the writer is trying to discover what they think.

Use AI after there is something to resist

Once an origin note exists, AI can be more useful because the writer has a position against which to compare suggestions.

I use it to:

• identify a counterexample that threatens the thesis;
• locate primary or authoritative source leads;
• compare two structures;
• find overlap with another Aethel essay;
• mark claims that require verification;
• expose paragraphs that sound polished but say nothing specific.

The crucial action is rejection. If every suggested section survives, the writer may be administering output rather than editing it.

A simple assistance ledger records three things: what changed, why it changed, and what evidence justified the change. It does not need to include every comma.

The authorship checkpoint

Before attaching a name, ask:

1. Can I state the contribution in one sentence without reading the article?
2. Can I identify the event, artifact, or reasoning that makes this piece specific to me?
3. Can I explain what every source supports and what it does not support?
4. Did I reject or substantially alter any generated recommendation?
5. Can I defend the important limitations without using boilerplate?
6. Would the article deserve to exist without search traffic or advertising?
7. Will I correct it if a reader produces better evidence?

A “no” does not prove that the text lacks value. It shows where responsibility is still thin.

Disclosure should describe roles

A useful public note can remain brief:

AI-assisted tools supported source discovery, structural comparison, and language editing. The author verified the cited material and approved the final argument and wording.

That statement should change when the process changes. If a system generated data, translated interviews, or created images, the disclosure may need to be more specific.

The note is trustworthy only when an internal process supports it. Repeating one universal sentence across every page can become another template.

The last draft written alone is not necessarily the final draft, and “alone” should not be treated as a purity test. It is the last artifact that lets the writer hear the question before the available answers become louder than it.

Keep it so the finished page has something to answer to.