llm-prompt-design

Put all prompt text and builder functions in a single module such as prompts.py or llm/prompts.ts, and call it from feature code instead of inlining prompt strings. This makes behavior changeable in one place and prevents drift across files.

Add an anti-slop rules block to every prompt: plain English, no hype, no clichéd closers like "I hope this helps!", no buzzword padding, and no invented facts. Keep it as a separate composable block so it applies consistently across features.

Pass a per-call context label that tells the prompt where it is being used, while keeping the role, structure, and rules blocks fixed. Vary only the context label rather than forking the entire prompt for each screen.

Degrade gracefully by preserving the user's original text instead of replacing it with an error string or blank output. Never hardcode API keys in prompt modules; load them through secure environment configuration.

Write simple assertions that the assembled prompt contains the required blocks: role/tone, output structure, anti-slop rules, and the context label. Spot-check output quality separately to catch drift from the expected shape.