Best Practices for BMAD

Practical guidance on team structure, estimating AI work, managing uncertainty, documentation standards, scaling AI development, and managing inference costs.

Team Structure for AI Projects

Effective AI teams blend traditional development skills with AI-specific expertise:

Estimating AI Work

AI tasks are inherently harder to estimate than traditional development. BMAD uses a modified estimation approach:

Feature: AI-powered ticket classification

Spike (1 day):
  - Can an LLM classify our ticket categories? YES
  - Baseline accuracy with zero-shot: 78%
  - With few-shot examples: 89%
  - Target accuracy: 90%

Estimates:
  Prompt engineering to reach 90%:  2-4 days
  Integration code:                 3 days
  Evaluation framework:             2 days
  Testing and QA:                   2 days
  Monitoring setup:                 1 day
  ─────────────────────────────────
  Total: 10-12 days (with buffer: 15 days)

Managing Uncertainty

AI projects have more unknowns than traditional software. BMAD manages this through:

1. Discovery Spikes

   Before committing to an AI approach, run a time-boxed spike (1-3 days) to validate feasibility. If the spike fails, you have saved weeks of wasted effort.

2. Progressive Commitment

   Start with the simplest possible AI approach. Only increase complexity if the simple approach does not meet quality targets. Many teams over-engineer when a well-crafted few-shot prompt would suffice.

3. Exit Criteria

   Define clear exit criteria for experiments: "If we cannot reach 85% accuracy in 3 days, we will try approach B." This prevents infinite optimization loops.

4. Fallback Plans

   Always have a non-AI fallback. If the AI component fails or does not meet quality targets, what is the manual or rule-based alternative?

Documentation Standards

BMAD projects require additional documentation beyond standard software docs:

• Prompt documentation: Every production prompt should have a README explaining its purpose, expected inputs/outputs, known limitations, and version history.
• Evaluation reports: Document evaluation results for each prompt version, including accuracy, sample outputs, and failure analysis.
• Model cards: For each AI model used, document its capabilities, limitations, cost, and any known biases.
• Runbooks: Create operational runbooks for common AI issues: quality degradation, model outages, cost spikes, and bias incidents.

Scaling AI Development

Cost Management

AI inference costs can grow rapidly. BMAD treats cost as a first-class engineering concern:

1. Model Tiering
   Use expensive models only when needed:
   - Simple tasks → Haiku ($0.25/1M tokens)
   - Standard tasks → Sonnet ($3/1M tokens)
   - Complex tasks → Opus ($15/1M tokens)

2. Response Caching
   Cache identical or similar requests:
   - Exact match cache: 90%+ hit rate typical
   - Semantic cache: 60-80% hit rate

3. Prompt Optimization
   Shorter prompts = lower costs:
   - Remove redundant instructions
   - Use concise few-shot examples
   - Limit output length with max_tokens

4. Batching
   Group requests when real-time is not needed:
   - Batch classification jobs
   - Process during off-peak hours

5. Budget Controls
   Set hard limits to prevent surprises:
   - Per-feature daily budget caps
   - Alert at 80% of budget
   - Auto-disable at 100% with fallback

Frequently Asked Questions