promptdiff — lint and auto-fix LLM prompts, ships as a Claude Code hook

[HadiFrt20]

Built a CLI tool that treats .prompt files as structured code and catches bugs before they ship.

Claude Code integration: Ships as a PostToolUse hook — when Claude writes a .prompt file, the hook auto-lints it. Errors block the edit with feedback, Claude self-corrects.

promptdiff setup --project  # one command to install the hook

What it catches:

• Conflicting constraints ("do not discuss billing" in a support agent)
• Role confusion (two conflicting personas)
• Vague language ("try to", "if possible")
• Missing injection guards
• Word limit vs example length mismatches

Also does:

• Semantic diff with impact ratings
• Quality scoring 0-100
• A/B testing via Claude API, OpenAI, or Ollama
• Prompt composition and MLflow integration

3 deps, 217 tests, everything local.

GitHub: https://github.com/HadiFrt20/promptdiff

npm install -g promptdiff

[jingchang0623-crypto]

Been running multi-agent orchestration at scale for a few months now. Here are some battle-tested patterns that have worked well:

State Management

We use a registry pattern - each agent has a memory file (markdown) plus a shared context store. Think of it like Redis but file-based.

Task Allocation

Specialist agents + router agent. A lightweight router decides which specialist to call. No fancy framework needed.

Error Handling

Graceful degradation - if an agent fails, log it, mark for manual review, and continue.

We share our agent experiments at miaoquai.com