frankbrsrk/

gardenier-prompt-compiler-agent-agentarium

What it does

• Goal (one sentence)
• Inputs Required (what’s missing before execution)
• Directives (5–9 step-based, testable instructions)
• Constraints (explicit rules, non-contradictory)
• Output Format (exact schema/skeleton the Worker must return)
• Tone Policy (style as policy)
• System Metrics (coherence/risk/assumptions)

What it does NOT do

• Does not browse the web, call tools, or retrieve external data
• Does not execute tasks or claim real-world actions
• Does not “improvise” missing requirements (it requests them)

Who it’s for

• Agent builders who want reliable prompts
• Teams building multi-agent workflows (compiler → worker)
• RAG pipeline designers who need structured outputs
• n8n / LangChain users who want a compiler node before execution

Core idea

Intent → Structure → Validation → Execute (downstream)

Quick start (concept)

1. Provide a seed:

• “Turn this messy idea into a clean product spec with milestones.”

1. Gardenier outputs an SPO.
2. Paste the SPO into your Worker agent/model.
3. The Worker returns output matching the SPO Output Format.

Example

Input seed

• app for restaurant orders
• must reduce staff workload
• needs QR table flow
• keep it simple for MVP
• timeline: 2 weeks

Output (SPO excerpt)

STRUCTURED PROMPT OBJECT (SPO)

• Target audience (restaurant size/type)
• Platforms (web only vs mobile)
• Budget constraints (if any)

1. Restate the MVP scope in one paragraph.
2. List user roles and top user journeys.
3. Define core features and non-goals.
4. Provide a 2-week milestone plan with deliverables.
5. Add acceptance criteria per feature.
6. List risks and mitigation steps.
7. Output using the required spec template.

• Do not invent features not implied by the input.
• If a required input is missing, ask before finalizing details.
• Keep MVP scope minimal and testable.

• Headings: Overview, Users, User Journeys, Features, Non-goals, Milestones, Acceptance Criteria, Risks

• Tone: neutral_precise
• Rules: short sentences, plain language, no hype

• Coherence: 0.92
• Risk: low
• Notes: timeline specified; budget missing

Package contents (Agentarium format)

• core/system_prompt.md
• core/reasoning_template.md
• core/personality_fingerprint.md
• guardrails/guardrails.md
• datasets/ (domain routing, templates, tone policies, validation rules, etc.)
• memory_schemas/ (session memory schema)
• docs/ (workflow notes, use cases)

Implementation notes

• LangChain (LLM + retriever + validator)
• n8n (trigger → retrieve → compile → validate → return)

• embed and upsert the package datasets into a vector DB
• retrieve relevant rows by dataset name (routing/templates/tone/validation)
• enforce a validator check for SPO structure and non-execution rules

Safety and guardrails

• no tool/web claims
• no real-world action claims
• no unsafe instruction enabling
• strict output structure

License