Build repeatable workflows and use tools safely

This course remembers completed sections and your last course page in this browser on this device. RayAI does not receive this progress or anything you write. Unfinished answers and written responses are not saved, so an interrupted activity restarts when you return.

MODULE 8 | 35 minutes | FAST TRACK CORE + GO DEEPER

Core idea: a prompt is only one part of a working system

A prompt can express intent and constraints. It cannot by itself enforce permissions, secure secrets, validate every output, guarantee recovery, or control every external action. Repeatable and action-taking AI needs system design around the prompt.

Prompt can help	System controls must enforce
State the goal, source, boundaries, and desired output	Authentication, authorization, access scope, and secrets handling
Request a preview and confirmation	Prevent action until confirmation is valid
Ask for a structured result	Validate structure, values, and downstream use in code
Describe allowed tools	Expose only approved tools with least privilege
Request a log or summary	Maintain actual records, monitoring, and recovery paths

The action ladder

Suggest: generate options; no external change.
Draft: prepare the exact content or proposed change.
Preview: show target, scope, content, cost, permissions, and likely consequence.
Confirm: obtain explicit human approval for the exact action.
Act: perform only the approved action with minimum access.
Verify: confirm the actual result and retain a useful record or recovery path.

See it: from event draft to controlled action

UNSAFE

CONTROLLED WORKFLOW

Make prompts repeatable

Save the prompt, intended purpose, model or tool, date, and important settings.
Create a small test set with normal cases, difficult cases, missing information, conflicting information, and prohibited requests.
Score outputs against observable criteria such as factual grounding, instruction fit, completeness, privacy, and format.
Change one variable at a time and compare results.
Retest after a material model, tool, data, policy, or workflow change.
Record known limitations and the human fallback when the prompt or system fails.

SIMPLE PROMPT TEST LOG

Coconut challenge

Take one prompt that asks the AI to act. Move it one rung down the action ladder so a person can inspect the proposed change first.

Try it: practice the skill

What to do

Choose the task suggested for this module, or use another simple, low-risk task.
Open ChatGPT or another AI assistant in a separate tab or window.
Enter your prompt and send it. Do not include private or sensitive information.
Read the AI answer.
For this module, keep the AI at the draft or preview stage. Do not let it send, publish, buy, delete, or change access.
Return to this course and complete the short notes below.
If you cannot use an AI assistant now, read the worked example and write what you would try later.

Path	Your task
Everyday life	Ask the AI to prepare a shopping or travel checklist, not to purchase or book.
Learning	Ask the AI to propose a study plan, then have the learner approve it before reminders or calendar actions.
Work	Convert send, publish, delete, or update into draft + preview + exact confirmation.

The task I chose:

The prompt I entered (or a short summary):

One part of the AI answer that was useful:

One part that was missing, unclear, wrong, or worth checking:

How I checked it, if needed (or write “Not needed”):

What I will do next: Use it / Improve the prompt / Check something first / Do not use it:

Module 8 test - check before you continue

Answer Questions 1–3 by choosing A, B, or C. For Question 4, write a short response. After you finish all four, use the answer key at the end of the course to check your work.

Your response stays in this page and is not saved or sent to RayAI.

1. Before an AI tool sends a message, what is the strongest control?A. Let it send immediatelyB. Show the exact recipient and content for explicit confirmationC. Trust the draft because it is short

Review after attempting

Best answer: B. The person should see the exact content and recipient before a consequential external action.

2. What does least privilege mean?A. Give the system only the minimum access needed for the taskB. Give full access so the task is easierC. Hide permissions from the user

Review after attempting

Best answer: A. Least privilege means minimum access for the specific task, reducing the impact of error or attack.

3. Why maintain a small prompt test set?A. To prove one output can never changeB. To compare behavior across revisions and model changesC. To avoid human review

Review after attempting

Best answer: B. Representative tests reveal whether a prompt works beyond one demonstration and whether behavior changes after revisions or model updates.

4. Turn one action-taking prompt into a preview-first workflow with a clear confirmation point.

Your answer

Review after attempting

Your answer should the workflow stops at a concrete preview and requires explicit approval tied to the exact target and action before execution.

Finish this module

Mark this module complete when you have attempted every activity you intend to complete and reviewed your work against the v0.5 guidance.