ohnobro398-png/quantum-skills

Quantum Prompt Engineer

Structure every prompt

• Give every serious prompt four parts: role, context, task and constraints. A prompt without constraints is a wish, not an instruction.

• Be specific about the output format: length, language, style and what to exclude. Specify the shape of the answer, not just the topic.

• State what the model must never do, not only what it should do. One clear negative rule beats five positive suggestions.

• For Claude, use XML tags to separate instructions, context and examples. For image and video models, use comma separated descriptive phrases instead of sentences.

Write generative prompts that work

• Build image and video prompts in this order: subject, action, setting, camera, lighting, style, quality tags. Every slot filled beats a long poetic sentence.

• End image prompts with quality and consistency tags, and keep a negative prompt list for what always goes wrong: extra limbs, distorted faces, watermark, text artifacts.

• Describe motion for video prompts as camera moves and subject moves separately. "Slow dolly in" plus "cat blinks" beats "cinematic video".

• For music prompts, name the genre, tempo, instruments, mood and structure. Reference songs guide better than adjectives.

• Keep proven prompt patterns in references/prompt-patterns.md and reuse the skeletons instead of rewriting from scratch.

Iterate scientifically

• Change one element per test: the subject, the style tag or the camera, never all three at once.

• Generate in small batches and compare side by side. Keep the winner, note why it won, and build the next test on it.

• When a prompt fails twice, rewrite the structure instead of adding more adjectives. Failure is usually missing constraints, not missing words.

• Save every winning prompt with its date and the model it worked on. Prompts are assets, not throwaway text.

Adapt the prompt to the model

• Different models read prompts differently. Test the same idea across models before declaring it good or bad.

• Short punchy prompts suit fast image models. Longer structured prompts suit reasoning models and video generation.

• When moving a prompt between models, keep the core description and retune only the style and quality tags.

• Never assume a prompt that worked last month works today. Models update silently, so re-verify hero prompts quarterly.

Available connectors

• Use muse.connector_search to find connectors and muse.skill_search to find related skills.

• If nothing here fits, look on the web for a maintained API or MCP server and set it up with Custom Connectors and MCPs, or build a custom connector.

Alternatives for better results

• Pair prompt work with the quantum_video_editor and quantum_ui_ux skills when prompts produce assets for videos or pages.

• For long prompt libraries, keep them as versioned files like the prompt pack, not as chat history.