How to Write Captions That Don't Sound Like Ads
A reusable skill for everyday product and personal posts. Ban hype. Structure every caption as one observation, one fact, one ask. Run it in Claude, ChatGPT, or Gemini without sounding like a media buyer.
Anyone posting a product or a personal update can get captions that don't sound like ads in one sitting: one observation, one fact, one ask, and a banned-word list. Run it in Claude, ChatGPT, or Gemini today, then cut anything you would not say to a friend at a table.
You do not need a brand voice workshop. You need a shape that forbids hype, plus a stop rule so the model cannot "punch it up." Punching it up is how a decent caption becomes a flyer.
This loop is a skill you reuse, not a prompt you retype. If you are still pasting "write me 10 captions" into a blank chat, read what a skill prompt actually is and save the card at the end.
A loop for captions that don't sound like ads
Ban the words first. Structure second. Variants last. If you reverse that order, every draft will arrive pre-hype.
1. Ban the hype before you draft
Paste this list into the chat and make it non-negotiable:
game-changer, revolutionary, must-have, obsessed, literally can't, wait for it, you need this, link in bio energy used as a sentence, 10x, unlock, elevate, seamless, delighted to announce, so excited to share.
Then add your own crimes. Every account has three. Write them down.
Instruct Claude, ChatGPT, or Gemini: "If a draft uses a banned word, rewrite the whole caption. Do not swap in a synonym. Change the thought."
Synonym swaps are how "game-changer" becomes "category-defining" and you still sound like a booth.
2. Feed one real moment, not a feature list
Your input is not the product spec. It is a moment.
- What you saw, heard, or broke.
- Who was there, if anyone.
- The one fact that would survive a skeptic.
- The one thing you actually want a person to do. Not "engage." A verb.
Example input: "Tuesday, 7:40 a.m. The pour-over tasted like the bag finally rested. 14-day rest printed on the label. Asking people what they rest their beans."
If you need somewhere to pull a saved moment later, keep a scrap pile in the Promptcrates library of reusable prompts. Do not mine it for slogans.
3. Force the three-line shape
Every caption, every platform, same bones:
- Observation — a concrete thing in the world. Time, place, sensory. No verdict.
- Fact — one claim you can defend. A number, a date, a material, a result.
- Ask — one question or one action a real person can do in under a minute.
Tell the model: "Three short paragraphs. Observation cannot contain the product name. Fact may contain it once. Ask cannot contain a discount, a countdown, or a superlative. 40 to 80 words. No emoji unless I used one in the input."
4. Generate three, pick one, do not blend
Ask for three variants that share the same fact and the same ask, but change only the observation. Different angle, same truth.
Pick one. Do not frankenstein a fourth from the best lines. Blended captions have no voice. They have a moodboard.
ChatGPT will offer a "punchier mix." Refuse it. Claude will offer a longer, more careful version. Shorten it. Gemini will offer speed. Speed is fine. Hype is not.
5. One human pass at speaking speed
Read the caption at the pace you talk. If you would not say it to one person, it is still an ad.
Kill:
- The fake we. You are one account.
- The fake urgency. If the ask is "reply with the city you are in," that is enough.
- The fake intimacy. "Hey friends" is a billboard in a hoodie.
Then stop. Do not add a fourth paragraph that "explains why it matters." If it mattered, the observation already showed it.
Save the loop with the Promptcrates tools for caption skills so next Tuesday you are not reinventing the ban list.
Reusable skill card
Trigger: You are about to post a product or personal update and the draft already sounds like a launch email.
Input: One real moment (time, place, sensory), one defensible fact, one ask, plus your banned-word list.
Output: Three captions, each 40–80 words, in the observation / fact / ask shape, product named at most once, zero banned words.
Stop: After you pick one variant and read it at speaking speed. No blending. No extra CTA paragraph. No hashtags in the skill output.
Takeaways
- Ban hype words before you generate. Synonym swaps still sound like ads.
- One observation, one fact, one ask. The product may appear once, in the fact.
- Generate three observations. Pick one. Never blend.
- Claude, ChatGPT, and Gemini all need a speaking-speed human pass.
- Save the ban list and the shape as a skill, not as a pinned "caption vibe."
Frequently Asked Questions
Can I still mention the product?
Yes. The product can be the fact. It cannot be the observation and it cannot be the ask. One mention. No slogan.
What if the platform wants a hook in the first line?
The observation is the hook. A specific thing you noticed beats "Wait for it" and "You need this" on every network.
Should I use Claude, ChatGPT, or Gemini for captions?
Claude is strict with banned words if you paste the list. ChatGPT is fast at variants. Gemini is fine for a first pass. All three need your human pass.
How many hashtags?
Zero in the draft. Add two after the skill stops, and only if they are searchable terms a person would type, not mood wallpaper.
What if I am posting about myself, not a product?
Same structure. Observation about the day. Fact you can stand behind. Ask that a friend could answer. Ego is just another hype word.