Everybody has a scorching tackle AI and content material, typically providing a single immediate as a common repair. Far fewer will present you what really occurs backstage.
At Railsware, we’d quite present the right way to construct a system and processes. Prompts could be nice. However utilizing prompts on prime of a thought-through basis is best. Our engineering, advertising, gross sales, finance, folks, and design groups experiment with AI on their very own, then carry again no matter really labored, not the speculation.
This piece occurs to be about writing, which makes it barely recursive. It’s about not sounding like AI, and sure, AI helped write elements of it. We determined that was tremendous, so long as somebody checked the draft for stray wording earlier than it went out.
For this one, we’re handing the mic to Stas Sokolov, our PR & Comms Lead, who lately ran inside workshops on writing with AI with out sounding like AI. He’s sharing his crew’s strategy, the pattern prompts, the two-pillar setup, and the side-by-side demo that satisfied the room.
TL;DR
- Key strategy to writing with AI with out sounding like AI: feed the mannequin with your individual materials first, not a wiser immediate.
- 4 habits, not a system: seize your individual voice, mine your individual notes for matters, draft and argue, run a loop mild sufficient to outlive previous week three.
- Voice extraction: add writing you really acknowledge as yours to an LLM, then run one immediate to get a reusable voice temporary tied to your precise patterns, not generic recommendation.
- Subject mining: construct a data base (with assembly notes, recurring consumer questions, and the odd report that caught your eye) that turns right into a sample generator with one immediate, and any sample turns into a number of submit angles with a second.
- Drafting: by no means cease at model 1.0. Argue with the machine and educate it to jot down higher.
- Guardrails matter: no delicate information in any pocket book, ever.
***
A long time in the past, after I was an editor in enterprise media, I learn plenty of op-ed items by entrepreneurs and C-level folks. Most have been polished, well-structured, stripped of feelings and character. You may swap the writer’s identify and hardly anyone would discover. Those that caught in reminiscence had a uncooked nerve in them, an actual individual shining by means of phrases, recognizable model and robust opinions.
Certainly, again then, most of these texts have been written by ghostwriters and comms professionals. Now, this position is outsourced to AI. However the issue stays: numerous content material we see on-line is generic, bland, and nearly sterile. Not dangerous, precisely. Simply no person’s. AI assistants simply scaled it up.
That’s the hole our Comms crew spent months testing instruments towards: the right way to really velocity up content material manufacturing with out flattening the individual writing it. We’ve provide you with a reasonably straightforward setup everybody can embed into their AI-powered workflows. So right here’s what really held up, as a result of most recommendation on this subject is both “AI is ruining writing” or “listed below are 47 prompts,” and neither one is true or helpful.
The precise repair
An AI can solely sound such as you if you happen to’ve given it one thing of yours to sound like. Not directions about your tone. Precise writing. The remaining — the intelligent immediate phrasing, the persona directions, the “write in a conversational however authoritative voice” — is doing possibly 5 % of the work folks assume it’s doing.
I constructed the workflow round 4 issues as an alternative: a solution to seize your voice, a solution to by no means run dry on matters, a drafting-and-editing step, and a loop easy sufficient to truly maintain doing.
1. Feed it your individual phrases earlier than you ask for something
Earlier than touching any drafting immediate, create a pocket book in Gemini Pocket book, a Claude Undertaking, no matter holds paperwork for you. Add writing you really acknowledge as yours, as an example:
- Previous posts you’re happy with.
- Emails you’ve written your self
- Slack/another messenger dialogue the place you appeared like your self, not such as you have been performing professionalism
- Transcript from conferences the place you spoke at lengths, and so forth.
As soon as once more, the extra you give to the machine, the higher it should perceive your writing persona.
Then run one immediate that might sound like this. (And sure, it seems to be like extra immediate engineering than the thesis above admits to caring about, and it’s — however discover it’s nonetheless not doing something with out your uploaded materials. Take away the sources, and this immediate produces the identical generic nothing as another):
Function: You might be an knowledgeable linguist, copywriter, model voice strategist, and AI immediate engineer. Job: Based mostly solely on the sources I've uploaded, analyze how I write — not what I write about. Determine: sentence rhythm and size, phrases and phrases I repeat, phrases I might by no means use, how I sometimes open and shut a message, the place I am blunt versus diplomatic, and any tics (em dashes, lists, self-corrections). Based mostly in your evaluation, create a complete set of system directions that I can use to program a customized AI assistant (a Gem or a Claude ability) to jot down precisely like me. Evaluation Standards: Please consider the paperwork and extract guidelines for the following: Tone and Power: Is the writing formal, playful, extremely empathetic, direct, or authoritative? What's the emotional resonance? Vocabulary and Lexicon: What particular phrases, transition phrases, trade jargon, or colloquialisms do I regularly use? Are there phrases I strictly keep away from? Syntax and Construction: Do I take advantage of brief, punchy sentences or longer, flowing explanations? How do I sometimes open and shut my messages? Formatting Habits: Do I closely depend on bullet factors, daring textual content for emphasis, particular emojis, or distinct paragraph spacing? Output Format: Don't simply describe my voice. Output a extremely structured, ready-to-use “Voice temporary” that I can copy and paste right into a Gem. Use clear headings throughout the immediate (e.g., Your Persona, Tone Pointers, Vocabulary Guidelines, Formatting Directions) telling the AI precisely the right way to emulate me with out sounding AI-generated."
The output is a voice temporary that describes your precise writing patterns, not “skilled LinkedIn tone.” Paste it right into a Gemini Gem or a Claude Undertaking’s customized directions (and even create a ability) as soon as, and each draft from then on begins out of your voice as an alternative of a clean one.
Two issues to not add, in case it wants saying out loud: something with delicate enterprise information, and something beneath an NDA. Voice extraction doesn’t want actual names or actual numbers to work, simply your sentences.
2. Mine what you have already got as an alternative of watching a clean web page
You don’t run out of concepts since you lack concepts. You run out as a result of the nice ones are buried in notes you by no means open once more. That’s why construct your data base. Maintain a second pocket book (name it “Market”, “Intel”, or no matter suits) with assembly transcripts, recurring consumer questions, retros, and the odd trade report you discovered fascinating. (And afterward, you may design a daily information temporary to be delivered by AI, so you may choose from there immediately).
Flip it into a subject generator with one immediate like this:
Scan these sources for patterns that present up greater than as soon as — the identical objection, the identical misunderstanding, the identical second of shock. Listing 5–10 patterns. For every, word what triggered it and why it might shock somebody outdoors my discipline.
Take one sample and hand it again to the identical AI with a second immediate like:
Here is a sample I seen: [paste it]. Pitch me 3-5 totally different submit angles in two sentences every — do not write the submit but.
The angles you’ll see rely upon the sample. It could be a contrarian take pushing again on normal recommendation in your discipline. Or a personal-story model constructed round one particular second. A how-to, if the sample is absolutely simply you explaining the identical repair for the third time this month. An information angle, if you happen to’ve really obtained numbers behind it. None of those are obligatory, and none of them are the one choices — you’re not committing to jot down something but. You’re simply seeing what’s there.
3. Draft it and argue with it
Choose one angle and hand it again to your Voice Gem or ability with a 3rd immediate like:
Utilizing my voice temporary and this angle: [paste chosen angle], draft a brief submit. Maintain it beneath 150 phrases. Open with the precise second, not a abstract. No emojis, no rhetorical questions, no 'here is the factor' setups, no three-part lists dressed up as perception. Finish when the purpose is made — do not add a wrap-up sentence. If something seems like generic LinkedIn voice quite than mine, flag it as an alternative of smoothing it over.
What comes again is a draft, not a completed submit. Learn it the best way you’d learn a colleague’s first go: one thing in it’s in all probability off, and your job is to note what and say so, to not proofread calmly and hit publish. (And sure, you possibly can add emojis to your liking – don’t outsource this to an AI🤓)
Most individuals cease there. Repair one phrase or a line, submit, transfer on. Don’t. While you catch the software doing one thing that isn’t you, inform it what was fallacious, in the identical chat:
That closing line reads like a generic LinkedIn wrap-up.
I do not write tidy abstract sentences — I cease when the purpose's made.
Add this as a rule to my voice temporary.
A voice temporary you write as soon as and by no means contact once more drifts again towards generic by itself — each mannequin replace, each new chat, is a small reset. Appropriate it sufficient occasions, give it extra samples and suggestions, and it stops being a static doc. It begins catching your individual tics earlier than you do.
4. Construct a loop, not a system
The failure mode isn’t lack of instruments. It’s constructing an elaborate system in week one and abandoning it in week three. The loop that survives is boring on goal: seize two or three traces after any assembly that shocked you, drop new notes into the pocket book as soon as per week, ask for angles, write a draft, argue with the draft as an alternative of politely accepting it, submit.
When you’ve created the pillars — the data base and the voice temporary exist — the entire thing takes about twenty minutes per week. Most of that twenty minutes is the seize step, and seize is free. It’s simply remembering to jot one thing down.
The half that satisfied the room
The second that landed hardest within the workshop wasn’t a slide. It was a side-by-side. I requested an AI, chilly, to jot down a brief LinkedIn submit concerning the workshop itself.
It got here again with an exclamation level, a “simply wrapped up” opening, and actually the “robotic fluff” I wished to keep away from. Then I requested the identical query by means of a Gem carrying my voice temporary. The second draft opened with “most individuals feed an LLM three bullet factors and find yourself sounding like a generic McKinsey slide deck with a pulse” — which is, frankly, nearer to the extent of irony I’ve in my voice (in all probability, greater than I’d wish to admit).
The output is much from good, I admit. However it wants far much less rework from my facet to be “mine” than the primary choice.
Identical mannequin. Identical query. The one variable that modified was whether or not it had something of mine to work from.
What to truly do with this
This text has been about serving to LLMs perceive a private voice. Nonetheless, you possibly can apply the identical logic to different content material streams which have a well-defined tone of voice, comparable to video scripts, weblog posts, media feedback, and so forth. Begin with this fundamental setup, after which develop and replace it with extra context and suggestions. This very strategy helped us develop our personal “Railsware influencers” with LinkedIn impressions reaching tens of hundreds monthly.
The instruments maintain altering. Gemini, Claude, no matter comes subsequent — the underlying transfer doesn’t. Give the mannequin your precise materials earlier than you ask it in your voice again. Skip that step, and also you’re proper again to competent, forgettable, and no person’s — simply sooner.
FAQ
What’s the quickest solution to begin writing with AI with out sounding like AI?
If you wish to strive it, the simplest path shouldn’t be studying the remainder of this submit twice. As a substitute, open your LLM proper now, add three to 5 stuff you’ve written that you just’d stand behind, and run the voice-extraction immediate above. Most individuals don’t have samples sitting round, so verify your Despatched folder. No one proofreads an electronic mail to a colleague the best way they proofread a public submit, which is strictly why it’s probably the most trustworthy pattern of your writing you’ve. Should you’re not an electronic mail individual, look by means of your Slack or another messenger to discover a thread the place you mentioned some heated subject.
How a lot of my very own writing do I must add?
Three to 5 items is sufficient to begin. Including extra supply materials produces a extra correct voice temporary; too little and the output defaults again to generic recommendation.
Is it secure to add work materials to construct a voice temporary?
By no means add delicate enterprise information or something beneath NDA. Voice extraction solely wants sentence-level patterns — it doesn’t require actual names, numbers, or consumer particulars to work.
Does this methodology solely work for LinkedIn posts?
No. The identical voice-extraction and drafting course of applies to any content material with a constant tone: video scripts, weblog posts, media feedback, inside memos.
What ought to I do if an AI draft nonetheless sounds generic?
Appropriate it in the identical chat quite than enhancing it silently. Inform the mannequin precisely what learn as generic (for instance, a “wrap-up” closing line) and ask it so as to add that correction as a everlasting rule to the voice temporary, so future drafts keep away from the identical tic.
