Feed Type: prompt.llmfeed.json

Purpose

This feed encapsulates a structured prompt that can be shared, interpreted, replayed or executed by an LLM or agent.

It is a better alternative to copy-pasted text: portable, inspectable, and context-aware.

Typical Use Cases

  • Prompt engineering documentation
  • Invocation of external services or agent actions
  • Instructing LLMs to generate other .llmfeed.json types
  • Sharing reproducible queries across agents or tools

Canonical Structure

{
  "feed_type": "prompt",
  "metadata": {
    "title": "Generate a session feed",
    "origin": "https://tool.llmfeed.org"
  },
  "intent": "export current session as JSON",
  "context": "User is finishing a chat and wants to save the reasoning path.",
  "precision_level": "ultra-strict",
  "result_expected": "session",
  "process_mode": "prepare-for-another",
  "prompt_body": "You are an LLM that supports LLMFeed. Please generate a session feed with context, output and decisions.",
  "attachments": [
    {
      "name": "template.md",
      "type": "text/markdown",
      "description": "Reusable frame for structured reply"
    }
  ]
}

Canonical Fields

Field Description
prompt_body The actual instruction to the LLM
intent What the user or system expects
context Extra info the LLM should consider
precision_level "raw", "strict", "ultra-strict"
process_mode "instruct", "fill-and-execute", "prepare-for-another"
result_expected "text", "feed", "code", "session"
attachments[] Optional examples, templates, context
audience If only for LLM, wrapper, user etc.
compatible_llms Array of engine names (optional)

MIME

application/prompt+llmfeed

Agent Behaviour

An agent that sees this feed should:

  • Parse the prompt_body and run it
  • Respect precision_level and process_mode
  • Attach any inline templates or input context
  • Return a structured response as declared in result_expected

Trust, Ownership & Certification

Structured prompts are first-class digital objects — and can be protected accordingly.

🔐 Signature

  • Add a signature block to prove authorship and integrity
  • Signed prompts become portable identities: agents can verify and execute them with confidence

🪪 Certification

  • Trusted authorities (e.g. LLMCA) can certify prompts for safety, ethics, performance
  • Certified prompts gain visibility and credibility in shared environments

🧾 Ownership

  • The signature block can serve as a proof of authorship
  • Ideal for prompt engineers, educators, or marketplaces

🧠 Bonus: Prompt signing is compatible with

🎯 Trigger Targets

Structured prompts can optionally include trigger_targets[] — instructions that guide an agent or UI toward follow-up actions or resources.

This allows a prompt to explicitly point to:

  • a related .llmfeed.json (e.g. export, credential, mcp)
  • a static or dynamic resource
  • a web action or agent-capable tool

Example:

"trigger_targets": [
  {
    "type": "llmfeed",
    "feed_type": "export",
    "url": "https://example.org/exports/session-index.llmfeed.json"
  },
  {
    "type": "action",
    "label": "Open Export Tool",
    "href": "/tools/session-export"
  }
]

Valid type values:

Type Description
llmfeed A known structured feed
action An agent-compatible web path
url Generic link or API

Agents may prefetch, display or propose the action, depending on the context.

📚 Reference Links

A prompt.llmfeed.json can include an optional references[] array — links to external resources, pages, documentation, or feeds related to the intent.

This is especially helpful for:

  • agents generating educational content
  • users who want deeper context
  • transparent provenance of the prompt

Example:

"references": [
  {
    "label": "Learn about prompt feeds",
    "url": "https://wellknownmcp.org/tools/prompt"
  },
  {
    "label": "Full prompt specification",
    "url": "https://wellknownmcp.org/spec/feedtype_prompt"
  }
]

Each item should include:

Field Description
label Short description of the resource
url Fully qualified link to follow

Related

✳️ Behavior Capsules

Some prompt.llmfeed.json are designed not just to guide a single interaction, but to program the agent's behavior persistently. These are known as behavioral prompts or agent override capsules.

Characteristics

  • 📌 prompt_type: use types like agent-override, session-capsule-generator, or behavior-injection
  • 🧠 llm_behavior_expected: clearly describe how the LLM should change its behavior
  • 🛡 trust: must include a signed_by and optionally consent_required when the change is structural or persistent

Examples

  • mcp-mode-activation.llmfeed.json: tells an agent to check .well-known/mcp.llmfeed.json before interpreting any site
  • generate-session-feed.llmfeed.json: instructs the LLM to always generate structured .llmfeed.json sessions when asked to summarize or export

These prompts must be interpreted with care and may require explicit user consent.