Traditionally, the workflow for creating enterprise content in the tech industry has gone something like this: A content strategist defines the brief—the purpose, the audience, the angle, the scope. A writer takes the brief and produces a draft. An editor reads the draft against the brief, checks the claims, catches the gaps, and sends it back. The writer revises. Stakeholders weigh in. The piece gets finalized.
In this workflow, each role has a distinct function. The strategist plans. The writer executes. The editor judges.
The execution part is now cheap
I spent many years as a freelance writer working with content strategists, product managers, and marketing leads. These were smart, educated people. While many of them were not writers per se, probably 80% of them could have generated a draft that got them good way toward the finish line.
But they didn’t have the time. My value prop for these busy people was basically a “shut the door” service—meaning that I had the luxury of shutting my door, focusing on the content, and generating a draft while my clients did their other work. Fast, clean execution was the service. It was valuable.
This service is worth less now. Not worthless, but less. AI executes faster and cheaper than any freelance writer. And what’s true for freelancers is true for in-house writers too.
The writer today still receives the brief. But invariably, the first draft comes from the AI, not from the writer’s own synthesis of the material. The writer’s job has shifted from producing the draft to managing the process that produces it.
Undeniably, something has been lost in this shift. When the writer’s primary job was producing the draft, they were forced to think through the material deeply enough to write it. This process was slow, sometimes frustrating. But it was also where judgment got built—where you might’ve noticed that the brief had a gap, an angle wasn’t quite right, or a claim didn’t hold up.
When the AI produces the draft, this process gets skipped.
The problem that doesn’t look like a problem
What AI does reliably is fill space convincingly.
Ask an AI to generate a draft and it’ll return something that looks complete. The structure is sound. The sections are all accounted for. The language is clean. The document presents itself as finished.
The problem is that the writer reviewing this draft is now doing the work of editing. Not that writers can’t edit their own stuff—but let’s be honest, writers and editors are different types.
The writer’s job, traditionally, was to jump into the fray. Take a brief, get lost in the material, make sense of the mess, and come out the other side with a draft. The writer was close to the work. Too close, sometimes. This is why a second set of eyes mattered.
The editor’s job was the opposite. Stand back. Read against the brief. Check the claims. Catch the gaps. Apply brand and compliance judgment. The editor’s value was distance—the ability to see the document as a reader would, not as its maker.
AI has put the squeeze on the jobs of writing and editing to the extent that it’s difficult to know where one begins and the other ends. For writers looking to maintain quality standards in spite of this squeeze, here are three practices worth building into AI-assisted workflows.
Keep your style guide in context
AI produces generic output when given generic context, and on-voice output when given specific voice context. The fix is straightforward: put your style guide in front of the AI before you start. In AI terms, everything you feed into a conversation—your instructions, your reference documents, your examples—lives in what’s called the context window. The more relevant material you load into it upfront, the more the model has to work with.
But context windows can get overloaded. If your full style guide is too long, convert it to a markdown file (a simple, lightweight text format that any AI tool can read easily) and tell the model which sections to prioritize. Brand voice and terminology are usually where the most important guidance lives.
Some ways to make this persistent:
- ChatGPT: Build a Custom GPT—a purpose-configured assistant with a persistent system prompt and uploaded reference files. Load your style guide once, and it’s available every time you open it.
- Gemini: Create a Gem—Google’s version of a purpose-configured AI assistant—with your style guide uploaded and instructions on how to apply it. Every conversation in this Gem inherits the context.
- Claude Projects: Create a project, upload your style guide, and add instructions telling Claude how to apply it. As with ChatGPT or Gemini, every conversation in the project inherits the context.
- Claude Cowork: Grant access to a local folder, store your style guide as a CLAUDE.md file, and Claude reads it automatically at the start of every session—no uploading required.
These practices raise the floor on first drafts, which means less time correcting the obvious and more time catching what actually matters.
Build in a verification layer
When source documents come with the brief, AI will draft fluently from the materials provided—and then keep going, filling gaps with plausible-sounding content that isn’t in the packet.
One approach I’ve used to combat this is a three-stage verification pipeline, each stage handled by a purpose-built AI assistant.
- The first extracts raw facts from the source documents into a ledger: every claim paired with its source, data gaps flagged, conflicts between documents noted. Nothing is summarized or interpreted—just extracted.
- The second drafts from the ledger only. It can’t reach outside the facts that were surfaced in step one.
- The third audits the draft against the ledger. Every claim in the output gets traced back to the source. Anything that can’t be placed gets flagged. The result is an integrity score and a list of unverified claims—each one a decision point for the writer.
This is a more involved setup than pasting documents into a chat window. But for high-stakes content where accuracy matters, it’s the difference between hoping and knowing that the AI stayed in bounds.
Run an adversarial audit
Once a draft exists, run it through a second LLM with an explicitly adversarial prompt. Something like: “You are a skeptical editor who suspects every claim and every number in this document. Find what’s unsupported, what’s imprecise, and what a critic would challenge. Return a numbered list of flagged claims, each with a brief note on why it’s problematic.”
This is different from the verification step above. The verification engine checks claims against your source documents. It knows what’s in the brief and the background documents. The adversarial audit has no sources. It’s questioning the draft on its own terms: logic, precision, and anything a skeptical reader might push back on.
The reason this works is the same reason the first model’s output is unreliable: LLMs are trained to be helpful, which means they’re inclined to agree and to fill gaps with plausible-sounding language. A second model prompted to distrust the first bypasses this tendency.
What the writer’s role is becoming
None of this replaces genuine writer and editorial judgment. Rather, these approaches create the conditions needed for judgment to operate at scale—distance, verification, structured challenge. The traditional division of labor between writer and editor achieved the same ends through different means. But now we’re in a different world.
When writers produced their own drafts, the slow, frustrating process of making sense of the material was where judgment developed. AI skips this process. The three practices above are an attempt to rebuild the scaffolding deliberately—to create the conditions for distance, verification, and structured challenge that the old workflow provided without anyone having to think about it.
← Back to Blog