builder-ai

Use before launching any LLM feature or when monthly API costs are growing unexpectedly. Requires token count measurement, call volume analysis, and cost projection at 10× scale. Blocks "it's cheap enough now" completions.

Use before shipping any LLM feature that touches users. Reviews prompt injection, hallucination risk, output misuse, agentic scope, and abuse vectors. Blocks "nobody will try that" completions.

Use when prompt cost is too high, latency is above threshold, or context window limits are being approached. Requires measurement before and after each reduction. Blocks "I shortened the prompt so it should be cheaper" completions.

Use before merging, deploying, or demo'ing any LLM feature. Requires documented eval results — pass rate, failure analysis, baseline comparison. Blocks "it looked good when I tested it" completions.

Use before merging any PR that adds an LLM API call. Every call must handle timeout, malformed output, low confidence, and refusal — with a defined, user-safe fallback for each. Blocks "add error handling later" completions.

Use when selecting a model for any production feature, or evaluating whether to switch models. Requires task-specific benchmarking — not leaderboard lookup. Blocks "GPT-4 is the best model" decisions.

Use whenever writing or modifying a prompt that will run in production. Enforces version-controlled prompts in prompts/<feature>/v<x.y.z>.md. Blocks "the prompt is in the code somewhere" completions.

Use when designing or auditing a retrieval-augmented generation pipeline. Requires data audit and query audit before any design decision. Blocks "I'll use the standard setup" completions.