Every generation tool has its own prompt syntax: Midjourney reads parameters after a double dash, ChatGPT has its own conventions, video models have a third set. Keeping all of that in your head is awkward. The fix is not to memorize the syntax but to set up a Claude Skill that asks which tool the prompt is for and returns it in the right format. Here is how to build such a skill from scratch.
The idea: a skill is a role that asks first and answers second
A Skill in Claude is a saved instruction you can invoke by command instead of explaining the task from scratch every time. A well-designed skill does not rush into writing the answer: first it clarifies what's missing (which model, which task), and only then returns the result in the right format.
Step 1 — find the Skills section
Go to Settings → Capabilities → Skills. The "Personal skills" section lives here — it works right away, with no connectors or integrations required. Hit "Create skill" to add a new one.
Step 2 — paste the prompt generator instruction
The skill instruction sets the role and the mandatory question before answering: which model the prompt is for and which task it serves. Here is the working text in full:
You are a prompt generator for AI models.
Ask: which AI (Midjourney, Flux, Sora,
or another) the prompt is for and what the task is
(photo, video, illustration). Do not answer
until you have both answers.
Then return the prompt in the format that this
specific model understands: with Midjourney
parameters (--ar, --style etc.) for Midjourney,
and a plain text description for models that
don't use parameters. Explain why you chose
that phrasing.
Save the skill under a clear name, prompt-generator for example.
Step 3 — invoke the skill in one line
From then on you don't paste the instruction again — only the task, with the skill hooked
in through /:
/prompt-generator
A prompt for a photoshoot by a lake,
for Midjourney, realistic style.
The skill will either return the prompt right away (if the task and the target model are clear from the request) or ask for what's missing — for instance, if the style or the aspect isn't specified.
The result: one skill, prompts for any task
Claude's output format adapts to the model you need, and you don't have to keep each tool's parameter syntax in your head. The same skill works equally well for a photoshoot prompt, a Reels cover or a post illustration — only the task wording in the request changes.
How to tune it to your niche
It's worth extending the base instruction with your own formats: if you regularly produce a specific type of visual (carousel covers, Reels frames, profile portraits), add it to the list of tasks inside the skill instruction — then the skill will recognize the format from a single word instead of a full description every time.
How we do it
The qvib.pro arsenal has cards both with ready skills and with a "how to build a skill from scratch" guide — for anyone who wants to assemble their own set for a specific niche. We use the same "clarify first, execute second" principle in the qvib engine when generating sites: the more precisely the inputs are pinned down before the start, the fewer edits after.
Related links
- The "Claude Skills" section of the arsenal: qvib.pro/arsenal/skills/
- The "Prompts" section of the arsenal: qvib.pro/arsenal/prompts/
- How the qvib engine works: qvib.pro/