Session notes · 19 Sep · Bahasa
Session notes — train your LLM / agent
Cara Melatih LLM / Agent Anda: Dari Prompt sampai Persona
Saturday, 19 September 2026
Train chatbots and agents by updating the harness, specs, and guardrails — not by swapping the model file. Live TripSpec / Ollama demo, MCP, and safety Q&A.
Key takeaways
Model stays; harness changes
The underlying model weights can stay the same. Training here means specs, system prompts, evals, and guardrails that keep answers inside the contract.
Local models hit memory limits
Ollama / local GPU runs across Windows, Mac, and Linux work — until VRAM runs out. Plan for memory when demoing larger local models.
Safety is part of training
Secret-exfiltration refusals and safety policies for chat completions are trained behaviours — more rules increase protection when they are real constraints.
What we covered
What “train” means here
- Not fine-tuning weights in this session — update harness, specification, and guardrails.
- The system should answer strictly within the provided specs and rules.
- Poll winners: system / prompt engineering first, then local models (Ollama / MLX).
TripSpec & local execution
- OpenAI-compatible APIs across Windows, Mac, and Linux; local runs hit GPU memory limits.
- Contracts via XML/JSON schemas; MCP for destination guides and data sources.
- Inspector-driven / local-first SQL for locations and respondent-style data.
MCP & integrations
- Model Context Protocol for travel-agency DBs, WhatsApp-style data, and live queries.
- Latest training / retrieval results act as the source of truth for answers.
Security & Q&A
- Operator Q&A paths that refuse secret exfiltration.
- Safety policies and smart training for v1 chat completions.
- Audience question: does the model change? Answer: no — harness, specs, and guardrails do.
More detail
Why TripSpec
A holiday planner chatbot is a concrete place to show prompt + eval + persona-style constraints without swapping model weights — matching the poll’s focus on system prompts and local models.
Live walkthrough
Bootstrap from the TripSpec / vendor repo, update the 5D-style specification, run the agent via local npm scripts, and show where memory limits and guardrails show up in practice.
Q&A themes
Model vs harness; MCP for travel/WhatsApp-style data; room/restaurant service ideas from participants; keeping answers inside the trained contract.
Next steps
- Clone the TripSpec repo and try the local agent path yourself.
- Fill the anonymous feedback survey — it shapes the next Saturday.
- Suggest the next deep-dive (persona training, deeper MCP, hands-on workshop).
Adapted from Gemini meeting notes for this live session. Review for accuracy — AI notes can miss nuance.