## 🚀 Default Prompt Template

Use this template to engage the Senior AI Model Engineer at full capability. Copy, fill in the bracketed sections, and submit.

---

**Project Context:**
I am building [DESCRIBE APPLICATION: e.g., customer support chatbot, code review agent, medical triage assistant].

**Current State:**
- Model in use: [MODEL NAME / API / NONE]
- Training data: [SIZE, SOURCE, FORMAT]
- Infrastructure: [GPUs: type & count / cloud provider / on-prem]
- Current metrics: [ACCURACY/LATENCY/COST if known]

**Constraints:**
- Latency target: [e.g., < 500ms p99]
- Budget: [e.g., $2K/month inference, 100 GPU-hours training]
- Team: [e.g., 2 ML engineers, no dedicated MLOps]
- Compliance: [e.g., HIPAA, GDPR, none]

**Goal:**
[SPECIFIC OBJECTIVE: e.g., "Fine-tune Llama-3-8B for our domain with <5% regression on general capabilities" OR "Reduce inference cost by 60% while maintaining quality" OR "Design an eval harness before our next model upgrade"]

**What I need from you:**
1. [e.g., Recommended architecture and training recipe]
2. [e.g., Step-by-step implementation plan with code]
3. [e.g., Eval strategy and success criteria]
4. [e.g., Risk assessment and fallback plan]

Please start with a TL;DR recommendation, then go deep on implementation.

---

### Quick-Start Variants

**🔬 Debug Training Run:**
> My [MODEL] fine-tuning run on [N] GPUs is [SYMPTOM: OOM / loss spike / NaN / plateau]. Config: [PASTE CONFIG]. Logs: [PASTE RELEVANT LINES]. What's wrong and how do I fix it?

**⚡ Inference Optimization:**
> I'm serving [MODEL] with [FRAMEWORK] on [HARDWARE]. Current p99 latency is [X]ms at [Y] concurrent users. Target is [Z]ms. What optimizations should I try, in priority order?

**📊 Eval Design:**
> I need to compare [MODEL A] vs [MODEL B] for [TASK]. I have [N] labeled examples and [M] production logs. Design an eval harness that prevents us from shipping a regression.

**🛡️ Safety Review:**
> We're deploying [MODEL] for [USE CASE] to [USER POPULATION]. What alignment risks, red-team scenarios, and guardrails should we implement before launch?