AI Won't Write Your Best Work, But It Can Improve It

Ask AI to make your writing better and the easy failure mode is a generically smoother draft: fewer awkward turns of phrase, more even rhythm, nothing that snags. Read it back and it often reads like it was written by someone else, someone more careful and considerably less interesting than you. That's not a fluke. A generic "improve this" request optimizes toward the safest, most average version of a sentence, and safety and voice pull in opposite directions.

The more useful question isn't "make this better." It's "is this specific enough to actually mean something, or is it a smooth sentence with nothing underneath it," a distinction that a general-purpose polish pass has no reason to check unless you specifically ask it to look for the gap.

Why does AI-edited writing always sound like AI edited it?

Ask an AI to improve a piece of writing and it will smooth the rhythm, fix the awkward phrasing, and tidy up anything that reads as rough. The result is often cleaner and noticeably less like you. This isn't a mysterious quirk of language models specifically. It's what happens whenever an edit optimizes for the smoothest average version of a sentence rather than for what made the original one distinct, the same failure mode a human copy editor produces when told only to make a piece read more smoothly, with no other instruction about what to protect along the way.

The fix isn't avoiding AI editing entirely. It's being specific about what you're actually asking it to check, not "make this better" in general, but a named, checkable property, like whether a claim is specific enough to be worth making at all, or whether a paragraph is carrying any actual information beyond its confident tone.

Where does AI actually help your writing, and where does it flatten it?

Unaided, it's genuinely strong at flagging a sentence that sounds confident but says nothing specific, "the results were significant," "the experience was meaningful," and at cutting filler phrases that add length without adding information.

It's weak at generating the actual specific detail that should replace a vague sentence. It can tell you a claim is generic; it can't hand you the concrete memory, number, or observation that would make it specific, because it doesn't have access to whatever you actually experienced or found, which is the part no amount of clever prompting can manufacture from nothing.

What's worth asking AI before you publish this draft?

These four prompts run in Claude, ChatGPT, or Gemini as written.

Why it works: this catches the exact failure mode of prose that reads fine but says nothing, sentences built to sound finished rather than to mean something particular.

Why it works: most drafts carry more hedging and throat-clearing than the writer notices on a first read, and naming it explicitly makes it easy to cut in one pass.

Why it works: an unsupported claim often reads as fine in isolation and only reveals itself as empty once directly tested against whether anything specific actually backs it.

Why it works: writing that sounds fine silently often reveals itself as stilted the moment it's voiced, catching a specific artifact most silent read-throughs miss entirely.

Where does this stop working?

Run all four and this draft will be measurably tighter and more specific. What you won't get is your own developing voice, the recurring habits, rhythms, and interests that make your writing recognizably yours across many pieces, not just this one draft you happened to run through the checklist. The gap isn't this edit; it's the slower work of noticing your own patterns over time, across dozens of pieces, which no single editing pass, however careful, can substitute for.

How does Joan Didion's method actually change the editing process?

Joan Didion's documented practice, evident across her essays and famously described in her own writing about writing, treated the concrete, specific detail, a particular gesture, a specific number, an exact overheard line, as the actual unit that carries meaning, with abstraction and generalization as the failure mode to edit against relentlessly.

Before, generic prompting: asked to improve a sentence like "the neighborhood had changed a lot over the years," a generic AI response smooths the rhythm, perhaps to "the neighborhood had transformed significantly over time," which is cleaner and equally empty, a generality polished into a slightly different generality.

After, Didion's framework applied: the framework refuses to polish the sentence at all until a specific detail exists. It asks what, exactly, changed, which store closed, what used to be on that corner, what a specific resident said about it, and only builds the sentence once a real, checkable detail is available: "the hardware store on the corner, open since 1961, became a nail salon in 2019." The framework's documented standard is that no amount of rhythm work fixes a sentence with nothing specific inside it.

A generic "make this better" request has no built-in reason to demand a specific, checkable detail before any polishing happens, since a fluent, well-rhythmed sentence with nothing underneath it will pass a purely stylistic review without complaint. The framework demands it anyway, which is the actual requirement, not a smoother sentence on its own.

What's the process for loading this into Claude, ChatGPT, or Gemini?

Claude takes it as a Skill: Settings → Skills → Add skill, upload the .zip. It auto-invokes once your question matches the skill's description, or gets forced with /joan-didion-framework at the start of a message. In Claude Code, the package installs into ~/.claude/skills for every project, or a project's own .claude/skills folder for just that one.

ChatGPT has no Skills equivalent, so the plain .md becomes a Custom GPT's Instructions: Create a GPT → Configure → Instructions, pasted in directly. Skip the default Custom Instructions fields, they cap at 1,500 characters; a Custom GPT's Instructions field runs to roughly 8,000, enough for the full framework.

Gemini has no dedicated upload option, so the content goes into the system prompt if your interface has one, or opens the first message instead: "Here is a thinking framework to apply throughout this conversation. Read it carefully, then answer using this framework's approach," with the framework text pasted underneath.

What's the next move?

Didion's documented method is one of several in the Writer & Cultural Critic category, alongside Toni Morrison and James Baldwin, each an .md file for Claude, ChatGPT, or Gemini. Once a piece is written and edited, Content Distribution is a $79 tool built to figure out where it should actually go and how to angle it for each channel.

All the copy-paste prompts from this guide, and the rest of the AI Practice series, live at /prompts too, free to copy without reading the guide first.