Intelligence Agent Playbook

This playbook targets power users and operators who want Tuff Intelligence to run repeatable multi-step execution flows, not one-off chats.

When to Use Agent Mode

• The task is multi-step (analyze → execute → validate).
• Tool calls are required (account, credits, settings, plugin actions).
• Risk approval is needed for destructive or sensitive operations.
• You need replayable traces for quality review.

Standard Operating Flow

1. Define objective in one sentence.
2. Set constraints for provider/model/tool scope.
3. Review plan (planner action list) before execution.
4. Approve risky actions before they run.
5. Run reflection and add follow-up actions when needed.

Provider & Prompt Practices

• Configure at least two providers for core capabilities (text.chat, content.extract) with fallback.
• Keep prompts as reusable templates bound to capabilities.
• Version prompts that change frequently (date or semantic version in title).

Recommended Templates

• Ops Triage: intent analysis → account/credits snapshot → risk list → next actions.
• Content Pipeline: input text → structured extraction → quality check → publish draft.
• Dev Collaboration: task decomposition → tool execution → failure convergence → rollback hints.

Common Pitfalls

• Single provider only: rate limits/timeouts break the full run.
• No approval gate for risky tools: increases accidental damage probability.
• Prompt not capability-bound: behavior drifts across entry points.

Pre-Launch Checklist

• Fallback provider is configured and tested.
• High-risk tools are approval-gated.
• Critical prompts are versioned and rollback-ready.
• Trace logs can pinpoint action/tool/provider.