Let’s start with the part most guides skip: this AI prompting process for writing might be a terrible idea for your work.
If you write literary fiction, personal essays, reported journalism, or any creative work where the authentic human voice is the entire point, you should probably stop reading here. Not because this guide will corrupt you, but because for those forms, the struggle to find your voice is the work.
Also on Mediabistro
You cannot outsource the struggle, the creative process, the indecision, and the messiness, and keep the art.
And you should also know this – even if you are writing B2B content, for example, using AI too much in your writing can be a huge issue. The AI and ranking systems themselves try their best to spot and reward human perspectives and opinions. So even in what you might call “less creative” fields such as brand content marketing, there are still commercial reasons why using AI prompts is a bad idea.
But, are you still here? Ok. Because there is a real and legitimate use case for what we’re about to discuss, and most guides either ignore the ethics entirely or drown them in so much hand-wringing that the practical value disappears. We’re going to do both: be honest about when this is a bad call, then try to be genuinely useful for the cases when it isn’t.
The Time and Place for AI
Writers wear a lot of hats. The same person who spends Tuesday writing a personal essay they’ve been carrying in their mind for three years also often has to produce:
- A weekly newsletter that goes out whether they feel inspired or not
- LinkedIn posts for a client who hired them to maintain a voice, not a byline
- Substack and newsletters that adhere to a fixed schedule
- Product copy for a brand they represent
- Blog content for a media company that runs on volume
- Email sequences, press releases, pitch templates, and the hundred other things that pay the bills
For that second category, writing that is fundamentally communicative rather than expressive, scaling your voice with AI assistance is a professional tool (and a good one if used correctly), and not a moral failing. The question is whether you’re doing it honestly and whether the output is actually good.
A ghostwriter who has spent years developing a client’s voice and now uses AI to help maintain that voice at scale is not necessarily “cheating,” unless the work itself suffers. A content strategist who trains a model on their own body of work to produce first drafts that they heavily edit is not cheating. A newsletter writer who uses AI to handle the 300 words of context-setting so they can focus on the 200 words of original insight is not cheating.
What is a problem is passing AI output off as deeply personal work, submitting AI-generated essays to literary journals, claiming AI-written journalism and facts as original reporting, or letting the model think for you rather than execute for you. That distinction matters. And really, if you are doing these things, you’re probably only hurting yourself, as “abusing” AI will tend to limit your own professional development.
What Google and AI Search Actually Reward (And Why This Changes the Calculus)
Here is something worth understanding before you go all-in or partially in on AI writing: the content landscape is not moving toward AI-generated text. It’s moving away from it.
Google’s E-E-A-T framework (Experience, Expertise, Authoritativeness, Trustworthiness) is explicitly designed to surface content written by people with real, lived experience in their subject. The “Experience” component added in 2022 specifically targets this: did a human actually do the thing they’re writing about? AI cannot demonstrate experience. It can describe it (often in a recursive manner). That is fundamentally different, and Google’s systems are getting better at telling them apart.
AI-powered search (Perplexity, ChatGPT Search, the AI Overviews in Google itself) pulls from sources that have demonstrated authority over time. Thin AI content farms are actively being deprioritized. The sites winning in AI-summarized results are the ones with genuine depth, original perspective, and the kind of specific, particular voice that signals a human actually wrote this.
That said, it may not exactly be the “human voice” that grants an edge in visibility, but rather providing originality. AI output is necessarily an average, and cites previously created data points. In other words, anything produced by AI has been said before in different ways, and exists on the Internet. It’s your job to create newness – and this is the only thing that will gain long-term visibility and distribution.
What does this mean for writers using AI? It means the human’s job is not being eliminated. The commodity is the AI-generated scaffold. The value is what you bring to it: the original observation, the counterintuitive take, the original data point or statistic, the original reported story, the specific anecdote from your career, the sentence construction that is distinctively yours.
If you train Claude on your voice and then let it write generic content in a pale approximation of that voice, you will likely not be rewarded with distribution. Even if it looks polished and “like your own,” there is still likely no originality.
If you train Claude on your voice, use it to produce structural drafts, and then infuse every output with the specific human observations that only you can supply, you may produce more original work than you could have written alone, faster than you could have written it.
The goal is not to replace your thinking, reporting, and experiential storytelling. It is to stop doing the parts of writing that don’t require your thinking.
What Claude Cowork Actually Is
Claude Cowork is Anthropic’s desktop application that allows you to create persistent, customizable AI workflows tied to your local files and folder system. Unlike the web version of Claude, Cowork can:
- Access folders on your computer directly
- Run custom “skills” you define in plain text files
- Maintain consistent behavior across sessions using those skill definitions
- Work with your actual documents without copy-pasting
A skill in Cowork is essentially a set of instructions you write once that permanently shapes how Claude behaves when you invoke it. For writers, this is the mechanism that makes everything below possible.
You are going to write a SKILL.md file that tells Claude exactly what your voice sounds like, what your best writing demonstrates, what patterns to use, and what AI-language patterns to purge. Claude will read this file every time you invoke the skill and use it as the foundation for everything it produces.
Note: It is helpful to save a chat as a favorite and link that chat specifically to a particular folder. Think of it this way – you’re not building a general-purpose “memory” in your Claude, you are working on a segmented project, governed by specific folders on your computer. Each project will have its own folder system and purpose.
Step 1: Install Claude Cowork
Download the Claude desktop app from claude.ai/download. Once installed, open it and look for Cowork mode in the settings. Enable it. You’ll be asked to select a working folder on your computer. This is the folder Claude will have access to for reading your writing samples and writing output files.
Create a dedicated folder for this project. Something like:
~/Documents/MyVoice/
Inside that folder, create two subfolders:
~/Documents/MyVoice/samples/
~/Documents/MyVoice/.claude/skills/voice/
The samples/ folder is where your best writing will live. The .claude/skills/voice/ folder is where your skill file will live. (The .claude directory is Cowork’s convention for finding skills.)
Step 2: Build Your Writing Voice Sample Library
This is the most important step and the one most guides rush. The quality of everything Claude produces depends entirely on the quality and volume of what you feed it. Bad samples produce a mediocre imitation of your work. Good samples produce something that actually sounds like you.
What makes a good voice sample for AI
You are not looking for your most popular pieces. You are looking for your most characteristic pieces. The writing that, if someone who knew your work read it anonymously, they would say “that’s definitely you.”
Good samples are:
- Unedited by a heavy hand: If an editor rewrote significant portions, those edits are not your voice. Use the version closest to what you submitted, not the published version.
- From your recent output: Your voice changes over time. Writing from five years ago may not represent how you write today. Use the last one to three years if possible.
- Representative of the type of writing you want to replicate: If you want to scale your newsletter, use newsletter issues as samples. If you want to scale your LinkedIn presence, use LinkedIn posts. The model will absorb the form as much as the style.
- Long enough to show patterns: A single paragraph tells Claude almost nothing. Aim for complete pieces: full articles, full newsletter issues, full posts. The pattern of how you open, develop an argument, use transitions, and close is as important as your word choices.
How many writing samples do you need?
Aim for 10 to 20 pieces that total at least 10,000 words. More is better up to a point; beyond 50,000 words you’re unlikely to get meaningful improvement, and you’ll be giving Claude more to process in every interaction.
That said, as a separate project, chat, and folder system, it might be a good idea to download your entire body of work, for example, if you run a blog or newsletter, and use that for surfacing interesting older content. But back to this project…
Save each piece as a plain text or Markdown file in your samples/ folder. Name them clearly:
samples/
newsletter-jan-2025.md
newsletter-feb-2025.md
blog-career-pivot-guide.md
linkedin-post-collection.md
feature-story-media-jobs.md
A note on your weakest writing
Do not include pieces you’re embarrassed by. Do not include pieces written under a deadline that don’t represent your best. Do not include pieces where you were writing for an audience that required you to dial back your voice. The model will average what it sees. If you include flat work, you will get even flatter imitations.
Step 3: Write Your SKILL.md File
This is the file that makes everything work and is currently what divides the average prompter from an advanced AI user. You’re going to save this Skill file in a folder, like:
~/Documents/MyVoice/.claude/skills/voice/SKILL.md
Below is a complete example skill template with explanations for each section. Read through and heavily edit the entire thing before using it, because the way you describe your voice to an AI is different from how you’d describe it to a human editor. When it’s done, you’re going to save it in a simple text editor as a file – and then grant your AI access to it.
Note: ensure that the folder paths such as /MyVoice and /Samples are correct in all areas inside this skill.md file and elsewhere.
SKILL.md Template
---
name: voice
description: Write in [YOUR NAME]'s voice. Use this skill any time content
needs to be written or drafted in their established style.
---
# [YOUR NAME] Voice Skill
## What this skill does
When invoked, this skill produces written content that matches [YOUR NAME]'s
established voice, based on their writing samples. It is not a general-purpose
writing assistant. It writes like them, not like Claude.
## Voice characteristics
[Write 4 to 8 specific, concrete sentences describing your voice.
Avoid vague terms like "conversational" or "engaging" — every writer
thinks they're conversational. Instead, describe the specific mechanics:]
Examples of what to write here:
- I use short paragraphs, rarely more than three sentences. The white space
is intentional.
- My sentences are direct. I rarely use subordinate clauses when a period
will do.
- I open with the most interesting thing first. I don't build to a point;
I start at the point and then support it.
- I use the second person "you" freely. I'm talking to someone, not at them.
- My humor is dry and appears once or twice per piece, usually buried rather
than announced.
- I reference specific numbers and details rather than generalizing. "Seven
years" instead of "years." "34%" instead of "most."
- I end pieces on action or implication, not summary. I don't restate
what I said.
## Reading my samples
Before writing anything, read all files in the samples/ directory of this
project. These are my actual published pieces. They are the primary source
of truth for my voice. If anything in this document conflicts with what
you see in the samples, trust the samples.
Pay special attention to:
- How I open pieces (my first sentence patterns)
- My paragraph length and rhythm
- How I handle transitions (or don't)
- My vocabulary range and the words I reach for
- How I use examples and specificity
- Where and how I use humor
- How I close
## What to produce
When asked to write something, produce:
1. A complete draft in my voice, ready for my review and editing
2. Nothing else — no explanations, no "here's what I did," no meta-commentary
I will edit the draft. Your job is to give me something worth editing.
## Banned language and patterns
Never use any of the following. These are AI-language patterns that will
make the output sound generic and not like me:
### Banned words (never appear in output):
- delve, delving
- navigate, navigating (unless literal navigation)
- landscape (as a metaphor)
- leverage (as a verb)
- utilize (use "use")
- foster
- holistic
- synergy, synergistic
- paradigm, paradigm shift
- robust
- streamline, streamlined
- cutting-edge
- game-changer, game-changing
- deep dive
- unpack (as a verb for ideas)
- unlock
- empower, empowering
- transformative
- innovative, innovation (unless quoting someone)
- seamlessly
- ecosystem (as a metaphor)
- journey (as a metaphor for any non-literal travel)
- framework (unless technical)
- actionable
- impactful
- bandwidth (unless literal)
- boilerplate
- circle back
- at the end of the day
- it goes without saying
- in today's world
- in today's fast-paced world
- the reality is
- the truth is (as a throat-clearing opener)
- make no mistake
- it's worth noting that
- needless to say
- in conclusion
- to summarize
- in summary
### Banned sentence constructions:
- Negative parallelism: "It's not about X, it's about Y" or
"Not X, but Y" constructions
- Em dashes used for dramatic effect or parenthetical emphasis.
Use a comma, a period, or parentheses instead.
- Sentences that begin with "It is important to note that"
- Sentences that begin with "It is worth mentioning that"
- Any sentence that opens with "Furthermore," "Moreover," "Additionally,"
"In conclusion," or similar formal transitions
- Lists introduced with "There are X ways to..." or "Here are X reasons..."
(use these sparingly and only when I explicitly ask for a list format)
- Filler intensifiers: "truly," "incredibly," "absolutely," "fundamentally,"
"essentially," "certainly," "undoubtedly," "remarkably"
- Rhetorical questions used as section openers ("Have you ever wondered...?")
- The word "very" before any adjective
### Structural patterns to avoid:
- The "X is important. Here's why." two-sentence opener
- Ending a section with a question to transition to the next section
- The three-example rule (AI defaults to three examples for everything;
vary this)
- Excessive hedging: "may," "might," "could," "arguably" used more than
once per paragraph
- Excessive and inauthentic emotionality about mundane subjects
- Summarizing what was just said before moving to the next point
## Quality check
Before outputting anything, re-read the draft and ask:
1. Does this sound like the samples, or does it sound like a capable
AI writing about the same topic?
2. Are there any banned words or patterns anywhere in the text?
3. Is every claim specific rather than general?
4. Are the paragraphs short enough?
5. Does it open with something interesting, or does it build to something?
6. Would a reader who knows my work recognize this as mine?
If any answer is "no" or "not really," revise before outputting.
Filling in the voice characteristics section
This section is where most writers get stuck, because describing your own voice is unexpectedly hard. Here is a process that helps:
Take your three best writing samples and ask yourself these questions for each one:
- What is the average number of words per sentence? (Count ten sentences and divide.)
- What is the average number of sentences per paragraph?
- What word does the first sentence of each piece do? (Does it state a fact, ask a question, make a claim, begin mid-action?)
- How many times do I use the first person “I” per 500 words?
- What are five words that appear in multiple pieces that aren’t common filler words?
- Is there a recurring structural move I make? (Starting with a story, ending with a call to action, using a concrete example to transition?)
The answers to these questions will be far more useful to Claude than abstract descriptions like “direct” or “approachable.” Claude does not know what your version of “direct” sounds like. But it can learn that you average 14 words per sentence, use “I” four times per 500 words, and always open with a specific fact rather than a generalization.
Step 4: Add Your Anti-AI Language Layer
The banned language list in the template above is a starting point. You need to personalize it with the patterns that specifically don’t sound like you.
Here is how to build your own extended list:
- Open Claude (without your skill) and ask it to write a 300-word post about something in your subject area, in a “clear, engaging professional voice.”
- Read the output and highlight every word or phrase that sounds slightly off, generic, or like something you would never write.
- Do this three times with different topics.
- Compile the patterns you found and add them to your banned list.
Common culprits that don’t make the standard lists but show up constantly:
- Starting sentences with “This means that…”
- “The key to X is Y” as a sentence structure
- Phrases like “at its core” or “at its heart”
- “What this looks like in practice…”
- “And that’s exactly why…”
- The phrase “more than ever” (as in “now more than ever”)
- Compound adjectives that sound like business-speak: “outcome-driven,” “value-added,” “results-oriented”
Also add any words that are simply not yours or you just don’t prefer. If you never use “whilst,” add it. If you never use “myriad” as an adjective, add it. The model does not know your vocabulary; it knows statistical patterns. You need to explicitly fence off the patterns that aren’t yours.
Step 5: Test, Iterate, and Calibrate
Once your skill is set up, run three or four test prompts before using it for real work. Ask it to write pieces similar to what’s in your samples: similar length, similar topic, similar format.
For each test output, do this:
- Read it cold, as if you’re a reader encountering your work for the first time.
- Circle every sentence that sounds wrong. Not just “AI-ish,” but specifically not like you.
- Look for patterns in what you circled. Is it sentence length? Transition choices? Paragraph structure? Something specific to your voice that you didn’t articulate in the skill file?
- Go back to the skill file and add that pattern, either as a banned construction or as additional voice guidance.
Plan on three rounds of iteration before the skill is tuned. Most writers find that the first pass captures about 60% of their voice, and each calibration round adds 10 to 15 percentage points. You will never get to 100%, and you shouldn’t try to. The final 20% is what your editing pass is for.
The editing pass is not optional
This cannot be stressed enough. The model produces a draft. You produce the work. The editing pass is where you:
- Add the specific observation that only you could have made
- Replace the representative example with the actual thing that happened to you
- Adjust the rhythm of any paragraph that still feels off
- Cut anything that the model added for structural completeness that you wouldn’t have included
- Insert the sentence you thought of while reading that makes the whole piece land
A 500-word Claude draft that you’ve edited for 20 minutes will be better than either a pure Claude draft or a piece you wrote entirely from scratch in 20 minutes. That is the actual value proposition.
Step 6: Running the Skill Day to Day
Once your skill file is in place and your samples folder is populated, here is the workflow:
- Open Claude Cowork and grant it access to your MyVoice folder
- In your message, invoke the skill by name: “Using the voice skill, write a 600-word newsletter section on [topic].”
- If Claude needs specific information, background, or a particular angle, include it in your prompt: “The key insight I want to make is [your actual original thought]. Build around that.”
- Review the draft with your editing eye
- Edit, add your specific observations, cut what doesn’t work
- Then, keep editing until you love it
- Publish the version that bears your fingerprints
The more specific you are in your prompts about the idea you want to express, the more useful the output. Vague prompts produce vague drafts. If you know what you want to say, tell the model what you want to say and let it handle the execution. If you don’t know what you want to say yet, that thinking is still your job.
Maintaining the Skill Over Time
Your voice will shift. What you were writing two years ago is probably somewhat different from what you’re writing now. Plan to update your voice samples every six months: add three or four new pieces, remove the oldest ones, and re-run your calibration tests.
Also update the skill file whenever you notice a new pattern that isn’t working: a new AI phrase that has crept in, a structural move the model keeps defaulting to that you don’t use, a characteristic of your newer writing that wasn’t present in your older samples.
The skill file is not a one-time setup. It is a living document that gets more accurate as you learn more about what makes your voice yours.
A Final Word on What This Is and Isn’t
If you train Claude on your ten best pieces and then let it write content you barely read before publishing, you haven’t scaled your voice. You’ve replaced it with a statistical approximation. The output will be “fine.” It will probably even pass a casual reader’s test. But it will be missing the thing that made those ten best pieces worth training on.
The writers who will get real value from this setup are the ones who understand that the model handles the architecture, and they still handle the life inside it. The scaffolding is Claude’s. The building is yours.
One key experience that you’ll notice is that your work output should increase – but it shouldn’t 10X. Writing with AI should still be hard, personal work. You’re striving to increase output of final results while also growing the quality. Most people who follow this process will notice that each piece of writing is still a significant lift. A sign that you’re doing it right might be that you notice a 25% improvement in speed – you don’t really want radical changes.
Where that line of time spent and effort sits is something only you can decide, and it probably sits in a different place for your newsletter than it does for your memoir, for your client work than for your bylined features, for the piece you’re producing on deadline than for the piece you’ve been carrying around for three years.
There is a time and a place. This guide exists to make you effective in the time and place where this is the right call. What that time and place actually is? That’s on you.
Topics:
Productivity
