Human-in-the-Loop Delegation Briefs — Plan, Draft, Verify

Cómo usar esta colección

## Mission This collection provides a safety-first workflow for delegating tasks to AI agents or human assistants. Instead of letting an agent act immediately, you use these prompts to generate a written contract (the brief) that defines exactly what the agent can do, what it must ask, and where you must intervene. ## Prerequisites - A standard LLM chat interface that accepts long text pastes. - No terminal, API keys, or special software installation is required. - You must be willing to read the output carefully and fill in gaps marked as `[NEEDS MY INPUT]` or `MISSING`. ## Progression Steps 1. **Decision & Planning**: Start with the **Manual vs. Agent-Assisted Task Planner**. Define your task, inputs, constraints, and risk tolerance. This step produces a comparison table showing which steps you should do manually versus which are safe to delegate. It establishes the permission checkpoints you will need later. 2. **Low-Stakes Practice**: Use the **Restaurant Booking Brief** to practice the delegation format on a reversible, low-risk task. The prompt converts your notes into a constraints table, a booking script with fallback options, and a confirmation checklist. Note: This does not book anything; it prepares a script for a helper. 3. **Handling Sensitive Data**: Move to the **Draft-Only Survey Invitations** brief. This adds strict exclusion rules and an approval gate. It ensures that no recipient is silently dropped and that sensitive lists are handled with care. The agent only drafts the invitation; you must send it. 4. **File System Safety**: Apply the discipline to local files with the **Image Sorting Safety Brief**. This prompt requires you to define root folders, off-limits areas, and an undo plan before any file is moved. It defaults deletion to 'NO' and requires a dry-run report. 5. **High-Stakes Verification**: Finish with the **Form-Filling Test-and-Verify Brief**. This creates a sample-data test pack and a verification table for live web forms. It includes hard stop rules for submissions and passwords, ensuring the agent never submits real data without your explicit, manual review. ## Critical Limitations - **No Execution**: None of these prompts perform actions. They do not browse, send emails, move files, or submit forms. They produce documents. - **No Fact Verification**: The prompts cannot verify facts not provided in your input. If information is missing, they will label it as `MISSING` rather than guessing. - **Privacy Warning**: Do not paste sensitive personal data (like real account numbers or private addresses) into these prompts unless you are comfortable with it being stored in your chat history. Use placeholders instead.