The difference between an AI draft you delete and one you publish is rarely the model. It is the prompt. A vague request like “write 5 LinkedIn posts about our product” gives the model nothing to anchor on, so it returns category-average sentences that could belong to any competitor. This guide is about the specific prompt techniques that fix that, with example prompts you can adapt today.
The four parts of a prompt that works
Every strong social prompt carries four things. Drop one and the output drifts.
- Role and goal — who the model is writing as, and the single job of the post.
- Source material — the real page, claim, number or quote the post must be built on.
- Constraints — channel, length, format, banned phrases, CTA rules.
- Voice sample — two or three real sentences from your existing content so tone is copied, not invented.
A prompt missing source material produces confident nonsense. A prompt missing a voice sample produces correct-but-robotic copy. You need both.
Example: turning a feature into a LinkedIn post
Weak prompt:
Write a LinkedIn post about our scheduling feature.
Strong prompt:
You are a B2B content writer for a project management tool. Write one LinkedIn post (max 1,100 characters, no hashtags in the body) for operations managers at 50-200 person companies. Source: our new “dependency auto-shift” feature moves every downstream task automatically when one task slips, so plans stay realistic without manual rescheduling. Angle: the hidden cost of manually updating Gantt charts after every delay. Open with a specific frustration, not a definition. End with a soft question, no link. Match this voice: “We’re not here to add another dashboard. We’re here to delete the 30 minutes you waste every Monday.”
The second prompt removes almost every degree of freedom that produces filler. It names the reader, the format limit, the proof, the opening style and the tone. That is why specificity beats volume in prompting.
Reusable prompt patterns
These patterns work across channels. Swap the bracketed variables.
Objection-to-post: “Take this common objection from our sales calls: [objection]. Write a [channel] post that acknowledges it honestly, then reframes it with this proof: [proof]. Do not sound defensive.”
FAQ-to-thread: “Here is a question customers ask: [question]. Write an X thread of 4-6 posts answering it. First post is the hook with a counterintuitive claim. Each following post adds one concrete point. No thread-bait phrases like ’let’s dive in’.”
Proof-first: “Write 3 Instagram caption options built around this result: [metric/outcome]. Lead with the number. Keep each under 125 characters before the fold. Different angle per caption: speed, cost, peace of mind.”
Repurpose: “Compress this blog section into one [channel] post. Keep the strongest single idea, cut the rest. Match the reading level of someone scrolling on a phone.” This pairs well with a full content repurposing workflow .
Constraints that raise quality
Negative constraints often do more than positive ones. Telling the model what to avoid removes its most common crutches.
| Constraint | What it prevents |
|---|---|
| “No phrases like ‘in today’s fast-paced world’” | Generic AI openers |
| “Max one adjective per sentence” | Hype stacking |
| “Cite the specific number, never ‘many’ or ‘most’” | Vague claims |
| “No call to action except a question” | Pushy CTAs that hurt reach |
| “Use the word the customer would use, not our internal term” | Jargon |
Keep a running list of these in your own notes. Most teams find five or six constraints handle 80% of the editing they used to do by hand.
Add the source, not just the topic
The single biggest upgrade is pasting the actual source text into the prompt instead of describing it. A pricing page, a support ticket, a case study paragraph, a five-star review. The model then works from your real facts rather than its training-data average. This is also why website-to-social systems are useful: they pull the source automatically so every draft is grounded. Pages like FAQs and pricing are unusually rich, covered in social media from FAQs and social media from pricing pages .
Iterate in two steps, not ten
Good prompting is rarely one shot. The reliable loop is short:
- First pass: generate three angles, not three finished posts. Pick the angle that is most specific to you.
- Second pass: ask for the full draft on the chosen angle, with the voice sample attached.
Asking for finished posts up front wastes effort polishing copy you will throw away. Choosing the angle first keeps the work cheap.
A prompt is only the first stage. Once a draft exists it still needs a human pass before it goes live, which is why prompting sits alongside AI posts with human review . Strong prompts also depend on a documented brand voice for AI social media so the voice sample is consistent across writers.
Utin is being built so these prompts run automatically from a scan of your website, with the source and voice already attached. If you are tired of writing the same setup into every prompt, you can register interest in the early pilot.