For the past fourteen months, roughly half of my working week has been spent inside AI writing tools. I have drafted product pages, newsletters, comparison posts, and one very stubborn white paper across Claude, ChatGPT, and Gemini, and I kept a log of every brief I sent and every draft that came back. The pattern in that log is blunt: the tool almost never decided whether the output was generic. My brief did.
The turning point came on a client project in late 2025. I sent the same article request twice, ten minutes apart. The first prompt was two sentences. The second was the same two sentences plus a paragraph about the reader, three snippets of my own writing, and a short list of phrases I never want to see. The first draft could have been published on any of ten thousand blogs. The second one sounded like something I would actually sign. Same model, same day, same topic. The only variable was context.
This article is the briefing system I built out of that log. It is not theory. Every layer below exists because leaving it out produced a flat draft I had to rewrite.
Why AI Writing Turns Generic in the First Place
A language model predicts the most statistically likely continuation of your prompt. When your prompt could have been written by anyone, the safest continuation is text that would satisfy anyone, which is exactly what generic content is. Three gaps cause most of it:
| Gap | What happens in the draft |
|---|---|
| 01 No defined reader | Without a specific audience, the model writes for an average of all possible readers, so it explains basics to experts and hedges every claim. |
| 02 No private information | If everything in your prompt is publicly known, the model can only remix publicly known ideas. Original inputs are the only path to original outputs. |
| 03 No voice reference | With no writing sample to imitate, the model defaults to its house style: tidy, symmetrical, agreeable, and identical to everyone else's drafts. |
The Four-Layer Context Stack
After enough failed drafts, I standardized my briefs into six layers. I keep them in a text file and fill them in before the first prompt goes out. When a draft comes back generic, I can almost always trace it to a layer I skipped.
| Layer | What it contains | What breaks without it |
|---|---|---|
| 1. Reader | Who they are, what they already know, what they are trying to do this week | Wrong depth, obvious explanations, hedged advice |
| 2. Job | The single action or belief the piece should produce | A summary of the topic instead of an argument |
| 3. Voice | Two or three samples of writing you want it to sound like | Default AI cadence and vocabulary |
| 4. Raw material | Your data, quotes, screenshots, test results, customer language | Recycled common knowledge |
| 5. Constraints | Length, structure, format, reading level, what to include | Bloated intros and padded sections |
| 6. Anti-brief | Phrases, claims, and moves the draft must never make | AI tells and clichés survive to publication |
Layer 1: Describe the Reader Like You Have Met Them
"Write for marketers" is not audience context. It narrows nothing, because marketers range from interns scheduling social posts to CMOs approving seven-figure budgets. The test I use: could the model make a decision with this description? If not, it is decoration.
PASTE-READY EXAMPLE Reader: a founder of a 4-person B2B SaaS company doing about $30k MRR. She writes the blog herself on Friday afternoons. She already knows what SEO is and does not need it defined. Her real problem: posts take her 6 hours each and still read flat. She is skeptical of AI content because she has seen competitors publish obvious slop. Assume intelligence, skip the basics, and never talk down to her. |
That paragraph took me four minutes to write and it changes everything downstream: vocabulary, depth, which objections get addressed, even sentence length. In my log, briefs with a reader paragraph of this quality needed about half as many revision rounds as briefs without one.
Layer 2: Voice Samples Beat Voice Adjectives
The most common voice instruction I see is a string of adjectives: "conversational, professional, engaging." Models cannot do much with adjectives because every writer defines them differently. What they can do, remarkably well, is imitate a concrete sample. Paste 200 to 400 words of writing you love, yours or a publication you admire, and label it clearly as a style reference rather than source material.
| Approach | Instruction | Result in my testing |
|---|---|---|
| Weak | "Write in a friendly but professional tone" | Same default cadence as no instruction at all |
| Weak | "Sound human, not like AI" | Forced slang and fake casualness layered over the same structure |
| Strong | Paste 2 samples + "Match the rhythm and vocabulary of these. Note the short punchy sentences after long ones." | Recognizably close to the sample by the second draft |
| Strong | Samples + a named contrast: "Like sample A, never like sample B" | The best results I recorded, especially for cutting filler |
| Tip from my log: point out one specific mechanical habit in your sample, such as "I open sections with a claim, not a question" or "I use one-sentence paragraphs for emphasis." Naming the habit roughly doubles how reliably the model reproduces it. |
Layer 3: Feed It Something the Internet Does Not Have
This is the layer that separates publishable AI drafts from filler, and it is the one most people skip because it requires actual work. A model trained on public text can only give you public knowledge. If you want a piece that says something new, the new thing has to enter through your prompt. Before drafting, I collect whatever private material exists for the topic:
▪ Customer language: support tickets, sales call notes, review quotes, the exact words real people use to describe the problem
▪ Your numbers: internal benchmarks, campaign results, pricing history, anything measured rather than assumed
▪ Your opinions: a voice memo or messy bullet list of what you actually believe about the topic, including the unpopular parts
▪ Your experience: what you tried, what failed, what surprised you, with enough detail that no one else could have written it
A practical trick that has saved me hours: talk before you type. I record a three-minute voice memo answering "what do I know about this that a stranger would not," transcribe it, and paste the transcript into the brief with the note "these are my raw opinions, build the article's point of view from them." The drafts that come back are opinionated in my direction instead of neutrally balanced, and neutrally balanced is just generic wearing a suit.
Layer 4: The Anti-Brief, or Telling It What Not to Write
Negative context is underrated. Models respond well to explicit prohibitions, and a standing blocklist catches the tells that make readers close the tab. Mine has grown with every draft I have edited, and this is the core of it:
| Banned phrases | Banned structures | Banned moves |
|---|---|---|
| "In today's digital landscape," "game-changer," "unlock," "elevate," "delve," "seamless," "it's important to note," "at the end of the day" | Three-item sentence patterns repeated back to back, every section ending in a mini summary, rhetorical questions as section openers | Hedging every claim with "can" and "may," praising both sides of a tradeoff equally, restating the heading as the first sentence |
HOW I PHRASE IT IN THE BRIEF Hard rules: no phrase from the banned list. Commit to positions instead of hedging. If a sentence would survive in an article about a different product, cut it or make it specific to this one. Vary paragraph length. At most one rhetorical question in the entire piece. |
That last rule, "if a sentence would survive in an article about a different product, cut it," is the single most effective line in my template. It gives the model a test it can actually apply to its own sentences during generation.
A Real Before and After
Here is a condensed version of the two briefs from the client project I mentioned in the intro, so you can see the difference in raw material rather than in the abstract.
BRIEF A, WHICH PRODUCED THE GENERIC DRAFT Write a 1,500-word blog post about email deliverability for SaaS companies. Make it helpful and engaging with actionable tips. |
BRIEF B, WHICH PRODUCED THE KEEPER Same topic. Reader: technical founders who already warmed up their domain and still land in spam. They have read the top 10 Google results and found them useless. Our angle, from our own migration in March: the fix that worked was throttling sends to 200/hour for 3 weeks, and our open rate went from 31% to 47% (screenshots attached). Voice: match the two samples below. Structure: problem, our failed attempts, the fix, exact settings. Anti-brief: banned list attached, no "tips," no numbered listicle framing, commit to our position that list cleaning is overrated for this specific problem. |
The cost of a thin brief shows up here: a draft covered in corrections. Photo: sidewalk flying, CC BY 2.0
Brief B is maybe 130 words longer than Brief A. That is the entire cost of the difference between a draft I deleted and a draft that ranked. Context is cheap. Rewriting is not.
How Much Context Is Too Much?
More context is not always better, and I learned this by overcorrecting. For about a month my briefs ballooned past 2,000 words, and quality dropped, because the model started treating every stray note as a requirement and the drafts turned cluttered and obedient. The pattern that emerged from testing:
| Brief size | Typical result | When it fits |
|---|---|---|
| Under 50 words | Generic almost every time | Throwaway internal drafts only |
| 150 to 600 words | The sweet spot in my log | Most articles, emails, and landing pages |
| 600 to 1,500 words | Strong, if clearly organized under labeled headings | Long-form content with data and samples |
| Over 2,000 words | Instructions start conflicting and the draft turns cluttered | Split it: send background first, then the task |
The fix for large briefs is structure, not deletion. Label each layer with a heading, mark which parts are reference and which are instructions, and tell the model the priority order when rules collide. "If voice and length conflict, voice wins" has resolved more bad drafts for me than any clever phrasing.
The Pre-Prompt Checklist I Actually Use
□ Can the model make decisions from my reader description, or is it decoration?
□ Have I stated the one action or belief this piece exists to produce?
□ Are at least two voice samples pasted and labeled as style reference?
□ Is there at least one fact, number, or story in this brief that Google does not have?
□ Is the anti-brief attached, with banned phrases and banned moves?
□ If rules conflict, have I told it which one wins?
Six questions, roughly ten minutes. In my log, briefs that pass all six get accepted on the first or second draft over 80 percent of the time. Briefs that skip three or more almost always trigger a full rewrite, which costs more than the ten minutes ever did.
Final Verdict, After Fourteen Months of Briefing These ToolsI came into this expecting to learn prompt tricks, and I left with a slightly uncomfortable conclusion: the quality of my AI drafts is a mirror of how well I understand my own reader, my own voice, and my own point of view. Every time a draft came back hollow, the hollowness was already sitting in my brief. The model just amplified it. So my honest recommendation is this. Stop editing generic drafts into shape, because that is the expensive end of the pipeline. Spend the ten minutes up front instead: one real paragraph about the reader, two voice samples, one thing only you know, and a blocklist of the phrases you are sick of. That is the whole system. It is unglamorous, it feels like homework the first few times, and it is the only thing in fourteen months of logged experiments that consistently turned these tools from a slop machine into a genuinely useful drafting partner. The tools will keep improving. Your context is still the ceiling. |
