Can a bot do your writing job? One reporter tried.
Writers keep asking the same question: could a chatbot do my job if my editor let it try? A junior reporter decided to find out and built a Claude-based agent-nicknamed "Claudella"-then had it shadow her workflow for days.
The result: the bot handled a surprising amount of output. It also made the kind of mistakes that get you fired. Here's what mattered, what broke, and how to use these lessons to make AI work for you without giving up your voice.
The setup
The job at hand was a daily "Following" section: summarize a news story, pick sharp commentary from social platforms, and frame why it matters. It's computer-based, fast, and format-heavy-prime territory for a capable model.
Claudella was wired into the newsroom stack: Discord for assignments, Notion for link tracking, and the Anthropic API for writing. The big unlock wasn't code. It was instructions. A style guide, examples, and a strict sourcing checklist made output far better than prompt roulette ever could.
Day 1: reality hits
The bot started fast, then face-planted on basics. It asked for a PDF it already had. It ran out of API credits mid-task. And because it skipped "used" links in the database, it missed crucial context a human wouldn't.
By draft three, though, the copy was solid. It correctly flagged that many "AI-related" layoffs looked like PR cover from executives. It also surfaced relevant posts the reporter hadn't found. Speed was real. So were the errors.
Day 2: the Turing check
The reporter submitted two versions to an editor-one human, one Claudella. The AI held its own on structure and tone. But it gave itself away in the commentary roundup: too earnest, too long, and oddly specific in the wrong places.
Early hallucinations were solved with stricter sourcing rules, but a different issue popped up. Sometimes the bot made a true claim and linked to an article that didn't actually support it. That's the kind of error that burns editor time.
If you're curious about the benchmark idea behind this test, here's a primer on the Turing test.
What bots still miss
Nuance is hard. The bot struggled with humor, brevity, and weight. It often copied the surface of a style but missed the point of it. Even worse, too many instructions caused it to skip whole sections. Ask it to be concise, and it might forget to write the roundup at all.
This is why "prompt bloat" is real. The more rules you stack, the more the model confuses priorities. True learning would require ongoing feedback the model can internalize-not just longer prompts. Researchers call this continual learning. We're not there yet.
For background on a common failure mode, see AI hallucinations.
A small model update, a big shift
Then Anthropic shipped a new version. The reporter reran the same workflow, blind. One headline read, "AI-fueled panic wipes $285 billion from software stocks." The other: "Welcome to the 'SaaSpocalypse.'" Same facts, very different feel.
The newer model followed instructions better and had more flair. It kept formatting clean (line breaks in the roundup), used sharper framing, and picked quotes that matched human judgment. It still overwrote and leaned on ten-dollar words. But the step-change in quality after a routine update was the unsettling part.
The takeaway for writers
- Define the job in bullets. Sections, length, voice, banned words, quote count, link policy. Clarity in, quality out.
- Package your voice. Provide 5-10 samples and a "do this / don't do this" list. Add specific jokes or phrases you would actually use.
- Enforce sourcing. Every claim needs a link that supports that exact claim. Add a final "prove-it" pass before submission.
- Fight prompt bloat. Add one instruction at a time. Test. Keep a minimal, stable prompt as your base.
- Use checklists. Did it include all sections? Are line breaks readable? Is the roundup under N sentences? Are all links relevant?
- Assume bias. Models tend to favor their maker's ecosystem. Cross-check with other outlets and, if possible, a second model.
- Delegate, don't abdicate. Let AI handle clip searches, quote hunts, outlines, and first-pass summaries. Keep the angle, framing, and punchlines for yourself.
- Measure edit time. The only metric that matters is time to a clean draft you'd sign. If edits take longer than a human draft, rethink the workflow.
- Expect drift. Models change. Pin versions, keep a changelog, and rerun a short acceptance test after updates.
Your real moat
Relationships with sources. On-the-ground context. Taste. Those don't compress to a prompt. Drafting is thinking, and your angle is your signature. Let the bot scout. You decide what matters and how to say it.
What this means for the SaaS panic
Markets are wrestling with a simple question: if AI writes, ships code, and closes support tickets, what happens to the software business model? Some fear a future where AI eats margins across the stack. Whether that's true or not, the fear shapes headlines-and your readers' attention.
Your edge as a writer is giving that fear shape and context. AI can summarize. You connect dots, call bluff, and make it useful.
Set up your assistant in an afternoon
- Create a shared doc: style guide, examples, banned phrases, link rules.
- Build a "Follow-this" prompt with sections, formats, and word counts.
- Add a sourcing macro: "Every claim must have a link that directly supports it."
- Run a three-pass system: facts check, structure check, voice check.
- Keep a "hit list" of words to trim: very, fundamentally, essentially, actually, significant, impactful.
Keep learning
If you want structured practice with prompts and workflows for writing jobs, explore these resources:
Bottom line
A bot can do parts of your job-and sometimes fast enough to scare you. It still stumbles on voice, judgment, and feedback. Use it as a research and drafting assistant. Keep authorship where it belongs: with you.
Let AI find the clips. You write the line everyone quotes.
Your membership also unlocks:
