"Write some Instagram captions" — why the output is interchangeable
A social caption prompt that returns lines you could paste under any brand, and the rewritten version that does not.
The prompt, as typed
Write some Instagram captions for my business. Make them fun and engaging.
What it leaves the model to guess
- What the image or post actually shows
- How many captions, and how long each
- The brand voice, and what it never sounds like
- What the caption is trying to get someone to do
Graded before and after
Both versions below were scored by the same rubric the free prompt grader runs.
As typed
45/100 · grade D-
| Clarity of context | 5 |
|---|---|
| Structure | 3 |
| Specificity | 2 |
| Reusability | 0 |
| Decision logic | 0 |
Rewritten
90/100 · grade A-
| Clarity of context | 7 +2 |
|---|---|
| Structure | 6 +3 |
| Specificity | 8 +6 |
| Reusability | 6 +6 |
| Decision logic | 5 +5 |
The rewritten prompt
Write Instagram captions.
Post shows: {{post_subject}}
Brand: {{brand}} — voice is {{voice_description}}
Audience: {{audience}}
Goal of this post: {{goal}}
Output 6 captions, each under 125 characters, each a different angle:
observation, question, contrarian take, behind-the-scenes, customer voice, plain statement.
Constraints: no emoji unless the brand voice line above asks for them, no
hashtag blocks, never open with "Introducing".
Rank the six by how specific they are to this exact post — a caption that
would work for any brand in the category ranks last. Why the rewrite works
The ranking rule is doing the real work here. "Would this caption work for a competitor?" is a test the model can apply, and it is the fastest way to push output away from the interchangeable middle.
Common questions
- Why does this prompt score badly?
- It scores 45 out of 100 (D-) because it leaves the model to decide 4 things you already know: what the image or post actually shows; how many captions, and how long each; the brand voice, and what it never sounds like; what the caption is trying to get someone to do.
- Does a longer prompt always score higher?
- No. The rubric rewards named inputs, a fixed output format and an explicit rule for choosing between options. Extra prose without those adds length and no score.
- Will this rewritten prompt work in ChatGPT and Claude?
- Yes. It is plain text with placeholders you fill in before running, so nothing in it is specific to one model or vendor.