# Model Prompting

> Prompt a specific model family well (Kling, GPT Image, Seedance…) once the endpoint is chosen, or when its outputs keep coming back generic.

- Skill: `kyh/model-prompting` (Agent Skill, multi-file: 4 files)
- Install (CLI): `npx skillmds@latest add kyh/model-prompting`
- Raw SKILL.md: https://api.skillmd.com/api/skills/kyh/model-prompting/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: kyh (https://skillmd.com/u/kyh)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/kyh/model-prompting

---


# Model Prompting

> **Runtime:** All endpoint calls use the `vg generate` CLI (`npm install -g vibedgames`, or `pnpm dogfood` in this repo). The API key lives on the vibedgames server, so there is no per-machine setup. See the `generate` skill for the command reference.

Model families have meaningful prompting nuances. A prompt that works on GPT Image 2 (long, structured, exact text in quotes) will fail on Happy Horse (which wants ~20 plain-English words). This skill collects family-specific guides.

## When to load which reference

| Reference                                   | Load when                                                                                             |
| ------------------------------------------- | ----------------------------------------------------------------------------------------------------- |
| [kling.md](references/kling.md)             | Working with Kling video models (O3, v3), multi-prompt, element controls, Standard vs Pro tier        |
| [gpt-image-2.md](references/gpt-image-2.md) | Working with `openai/gpt-image-2` or `/edit`, structured prompts, EXACT TEXT, multi-image compositing |
| [happy-horse.md](references/happy-horse.md) | Working with `alibaba/happy-horse/text-to-video` or `image-to-video`, brevity-first, camera language  |

## Universal principles (all families)

1. **Visual facts beat prestige adjectives.** Replace "stunning, cinematic, masterpiece" with "overcast daylight, brushed aluminum, 50mm feel."
2. **Style tags need visual targets.** "Minimalist brutalist" → "cream background, heavy black sans serif, asymmetrical type block, generous negative space."
3. **One controlled variable per iteration.** Comparison only works when one axis changes at a time.
4. **Inspect schema before assuming a control exists.** `vg generate schema <endpoint_id> --json`. Negative prompt, seed, multi-prompt, and reference-image fields differ across families.
5. **Per-family rules override universal advice.** Happy Horse rejects what GPT Image 2 rewards.

## Catalog cross-reference

For "which model do I use" questions, see `model-catalog`:

- Text-to-image endpoint selection → [model-catalog/references/text-to-image.md](../model-catalog/references/text-to-image.md)
- Text-to-video endpoint selection → [model-catalog/references/text-to-video.md](../model-catalog/references/text-to-video.md)
- Image-to-video endpoint selection → [model-catalog/references/image-to-video.md](../model-catalog/references/image-to-video.md)