Course resource
Writing in Your Own Voice
How to get AI-assisted writing that reads like you wrote it — because you did, with help.
A note on framing: the goal here is not evading detection. It is that generic writing is worse writing. Text that could have been written by anyone persuades nobody, and readers have become very good at noticing. The disclosure question is covered in Week 8; this is about quality.
Why AI writing sounds the way it does
Three causes, each with a different fix.
1. It writes to the middle. Trained on everything, it produces the average. The average is inoffensive and forgettable. Fix: give it your specifics.
2. It hedges structurally. "It's important to note", "while X, it's also true that Y". Fix: forbid the constructions.
3. It has no experiences. It cannot tell me what happened to you last Tuesday. Fix: you supply the specifics; it shapes them.
Build a voice profile
Twenty minutes once, useful for years.
Take three things you wrote that sound most like you. Then:
Here are three pieces I wrote. Analyse my voice, not my topics.
Tell me:
1. My average sentence length, and how much it varies
2. Words and constructions I use repeatedly
3. How I open and how I close
4. My register — formal to casual, on a 1-5 scale
5. How I handle uncertainty: do I hedge, state flatly, or ask?
6. Whether I use humour, and what kind
7. Three things that are distinctive about how I write
Then write a "voice profile" I can paste into future prompts.
Be specific. "Conversational" is useless.
[PASTE THREE SAMPLES]
Edit what comes back — it will get some of it wrong — and save it.
Your voice profile
| Sentence length | |
| Rhythm (uniform / varied) | |
| Formality 1–5 | |
| Words I actually use | |
| Words I never use | |
| How I open | |
| How I close | |
| Humour | |
| How I handle uncertainty |
The banned list
These are the words and constructions that mark text as machine-written to a 2026 reader. Paste into every writing prompt:
Never use: delve, leverage (as a verb), robust, seamless, landscape,
tapestry, realm, myriad, plethora, testament, "navigate the complexities",
"in today's fast-paced world", "it's important to note", "it's worth noting",
"that being said", "at the end of the day", "unlock", "elevate", "game-changer".
Never use the construction "It's not just X — it's Y."
Never open by restating my question.
Never close with a paragraph that summarises what you just said.
Never use an em dash to create false drama.
The list matters less than the habit: when you notice a phrase appearing in everything, add it.
The workflow that actually works
Do not ask it to write, then edit what comes back. You end up editing generic text toward your voice, which is slower and worse than the alternative.
Do this:
1. You supply the substance. Voice-note or bullet your actual thinking. Messy is fine.
Here are my raw notes. They are disorganised.
Do not write anything yet. Tell me what the actual argument is,
and what is missing.
[NOTES]
2. Agree the structure.
Propose three possible structures. For each, say what it emphasises
and what it buries.
3. Then draft, with the voice profile.
Write it using structure [N] and this voice profile: [PROFILE]
Keep every specific detail from my notes. Do not add examples I did not give you.
[BANNED LIST]
4. Put yourself back in.
This is the step that matters and the one people skip. Go through and add:
- One thing only you know — a number from your own work, something a client said, what went wrong last time
- One opinion the average person in your field would argue with
- One sentence that sounds like speech rather than prose
If you cannot add any of these, the piece has nothing in it and no amount of editing will fix that.
The read-aloud test
Read the finished piece out loud. Mark every place you:
- stumble (the sentence is badly built)
- get bored (cut it)
- sound like someone else (rewrite in your words)
- would not say this to the person's face (soften it or mean it)
This catches more than any detector, and it catches the things that actually matter.
Diagnosing a flat draft
| Symptom | Cause | Fix |
|---|---|---|
| Could apply to anyone | No specifics | Add your own data, names, incidents |
| Every sentence the same length | Model's default rhythm | Break three sentences in half; join two |
| Nothing to disagree with | No position taken | State the thing you actually believe |
| Reads as a summary | Structure is a list, not an argument | Reorder so each paragraph earns the next |
| Confident but empty | Hedges removed, substance never added | Go back to step 1 |
The rule
If you would not put your name on a paragraph, do not. Not because a detector might flag it — because it is not good enough.