Idea: community-contributed prompt templates with quality scoring

[caishen-ai]

Cool project - love the focus on taste and style in AI outputs.\n\nIdea: What about a community-contributed prompt template gallery? Users submit prompts that produce high-taste outputs, and the community scores them. This would:\n1. Create a feedback loop for what 'good taste' actually means in practice\n2. Give new users a starting point beyond the defaults\n3. Surface edge cases where taste is subjective\n\nI've been collecting AI writing prompts that focus on producing non-generic, human-sounding content for my blog. For example, prompts that avoid the classic AI 'slop' patterns (delve/tapestry/crucial etc).\n\nIf this fits the roadmap, I'd love to contribute some examples. I also have a small collection of AI writing resources at caishen-ai.github.io/caishen-blog/order.html if it's useful.

[optimization2026]

I like this idea, but I think it needs structure to avoid becoming a random prompt dump.

A prompt gallery becomes much more useful if every submission carries enough metadata to explain:

what the prompt is for
which skill it pairs with
what visual direction it targets
what good output means
what failure modes it avoids
what examples prove it works

A minimal template format could look like:

{
  "id": "premium-saas-dashboard-dark-editorial",
  "title": "Premium SaaS dashboard with editorial contrast",
  "target_skill": "design-taste-frontend",
  "category": "frontend",
  "style_tags": ["saas", "dashboard", "dark", "editorial", "premium"],
  "prompt": "Design a premium analytics dashboard for...",
  "good_for": [
    "SaaS dashboards",
    "admin tools",
    "data-heavy product pages"
  ],
  "avoid_when": [
    "consumer landing page",
    "playful brand",
    "content-first blog"
  ],
  "expected_traits": [
    "strong information hierarchy",
    "non-generic card composition",
    "intentional typography scale",
    "clear visual rhythm",
    "usable contrast"
  ],
  "failure_modes": [
    "generic three-card grid",
    "random glassmorphism",
    "low-contrast muted text",
    "default AI purple gradient",
    "template-looking hero"
  ],
  "scoring": {
    "layout": 0,
    "typography": 0,
    "originality": 0,
    "accessibility": 0,
    "brand_fit": 0,
    "implementation_feasibility": 0
  }
}

This would make the gallery more like a taste benchmark than a prompt list.

A possible structure:

examples/templates/
├── frontend/
│   └── premium-saas-dashboard-dark-editorial/
│       ├── template.md
│       ├── metadata.json
│       ├── scorecard.md
│       └── example-output.png
├── brandkit/
├── imagegen/
└── redesign/

A machine-readable workflow could help here too.

For example, an optional AISOP companion could define how a community prompt is evaluated:

[
  {
    "role": "system",
    "content": {
      "protocol": "AISOP V1.0.0",
      "axiom_0": "Human_Sovereignty_and_Wellbeing",
      "id": "taste_prompt_template_evaluation",
      "name": "Taste Prompt Template Evaluation Workflow",
      "version": "1.0.0",
      "summary": "Machine-readable companion workflow for evaluating community-contributed prompt templates.",
      "flow_format": "mermaid",
      "loading_mode": "node",
      "tools": ["filesystem"],
      "params": {
        "template_path": "string",
        "metadata_path": "string",
        "output_artifact_path": "string?"
      },
      "system_prompt": "{system_prompt}"
    }
  },
  {
    "role": "user",
    "content": {
      "instruction": "RUN aisop.main",
      "user_input": "{user_input}",
      "aisop": {
        "main": "graph TD\n    load[Load template metadata] --> run[Run target skill with template]\n    run --> capture[Capture output artifact]\n    capture --> score[Score output quality]\n    score --> blocking{Blocking failures?}\n    blocking -->|yes| reject[Reject or request revision]\n    blocking -->|no| accept[Accept into gallery]\n    reject --> end_node((End))\n    accept --> end_node"
      },
      "functions": {
        "load": {
          "step1": "sys.io.read(template_path) -> template_prompt",
          "step2": "sys.io.read(metadata_path) -> template_metadata",
          "output_mapping": "template_submission",
          "constraints": [
            "Every template must declare target_skill, category, style_tags, good_for, avoid_when, expected_traits, and failure_modes.",
            "Templates without metadata should not enter the gallery."
          ]
        },
        "run": {
          "step1": "Run the declared target_skill using template_prompt.",
          "step2": "Capture model, agent, skill version, and runtime context when available.",
          "output_mapping": "template_run_result",
          "constraints": [
            "Do not score a template without at least one output artifact.",
            "Do not mix results from different target skills without recording it."
          ]
        },
        "capture": {
          "step1": "Capture screenshot, image, code output, or other artifact generated by the template.",
          "step2": "Record whether the output is frontend code, reference image, brand board, redesign patch, or writing sample.",
          "output_mapping": "output_artifact",
          "constraints": [
            "A prompt template should be judged from evidence, not just from how good the prompt sounds.",
            "If the output artifact is missing, mark the submission incomplete."
          ]
        },
        "score": {
          "step1": "Score the output on layout, typography, originality, accessibility, brand_fit, and implementation_feasibility.",
          "step2": "Identify failure modes from template_metadata.failure_modes that appeared in the output.",
          "output_mapping": "quality_scorecard",
          "constraints": [
            "Pretty but unreadable output should not score high.",
            "Generic AI layout patterns should reduce originality score.",
            "Implementation feasibility matters for code-oriented templates.",
            "Accessibility is part of quality, not an optional afterthought."
          ]
        },
        "reject": {
          "step1": "Reject or request revision when blocking failures exist.",
          "step2": "Return the reasons and the failed dimensions.",
          "output_mapping": "revision_request",
          "constraints": [
            "Do not accept templates that consistently produce slop patterns.",
            "Do not accept templates with strong visuals but severe readability failures."
          ]
        },
        "accept": {
          "step1": "Accept the template into the gallery with metadata, scorecard, and output evidence.",
          "step2": "Record recommended use cases and avoid_when cases.",
          "output_mapping": "accepted_template_record",
          "constraints": [
            "Accepted templates must include evidence and scorecard.",
            "Subjective taste edge cases should be documented rather than hidden."
          ]
        },
        "end_node": {
          "step1": "Return accepted_template_record or revision_request."
        }
      }
    }
  }
]

The important thing is that scoring should not reward only surface-level beauty.

A useful scorecard could include:

| Dimension | Question |
|---|---|
| Layout | Does the composition avoid default AI patterns? |
| Typography | Does the type system feel intentional? |
| Originality | Is there a clear visual point of view? |
| Accessibility | Is the output readable and usable? |
| Brand fit | Does it match the intended audience and tone? |
| Implementation feasibility | Can this be implemented without excessive complexity? |

This would help the gallery surface not only “good prompts”, but also prompt-context pairs:

this prompt works well for SaaS dashboards
this prompt fails for mobile
this prompt produces strong images but weak code
this prompt is beautiful but accessibility-poor
this prompt works only with image generation models

That would make it much more useful for new users than a flat community prompt list.