nextfrontierbuilds/moltbot-best-practices
Overview
This skill codifies best practices for AI agents to avoid common, costly mistakes. It teaches agents to confirm intent, show drafts before publishing, stop on command, and avoid unnecessary automation. The guidance is grounded in real failures and focuses on predictable, safe agent behavior.
How this skill works
The skill inspects agent workflows and interaction patterns, enforcing checkpoints like confirmation prompts, draft approvals, and post-action verification. It also provides failure-handling rules: time-box retries, escalate after repeated errors, and silence noisy failure narrations. The rules are lightweight and designed to be applied across Claude, Cursor, and GPT-style agents.
When to use it
- Onboarding new agents or agent configurations to reduce early mistakes
- Before enabling publishing or automation capabilities in an agent
- When agents interact with external tools (browsers, APIs, posting platforms)
- When users report unexpected automation or ignored stop requests
- To harden agents that manage multi-step or long-running tasks
Best practices
- Always confirm the user’s intent in a single sentence before starting work
- Show a draft and request explicit approval before publishing or posting
- Stop immediately when the user says STOP or requests "READ THE CHAT"
- Limit retries: 2–3 attempts or a 5-minute cap, then ask the user
- Prefer manual handoff over aggressive background automation for fragile tasks
- Process queued messages in order and re-read recent context before acting
Example use cases
- A content agent drafts a blog post, shows the draft, and waits for approval before publishing
- A support bot detects a flaky browser automation and offers a manual fallback to the user
- An agent receives multiple rapid messages; it reads all queued messages before responding
- A data-entry agent stops immediately when a user types STOP and confirms the session state
- A pipeline enforces 2 retries on an API call, then escalates to the user for guidance
FAQ
A clear affirmative from the user such as "Post it" or "Yes, publish"; implicit silence is not approval.
When should an agent spawn background workers?
Only when the user agrees to background processing after you explain estimated time and trade-offs.
How do I handle repeated tool failures?
Stop after 2–3 failures, report the issue concisely, and ask whether to retry, switch methods, or exit.
