NewAI Content Creation is now live in Early Access
Turning Point Training Instituteby Training Center
Skip to content
0%
Module 2 · Verify your environment and boundariesLesson 2 of 2
Video lesson

L04 · Set permissions, privacy and cost boundaries

IntermediateVideo lesson
6 min readIntermediate
Downloads & resources
Download this lesson's materials, ready to use.

Skill packs and prompts are provided in English so they work best with the AI tools. Use them as-is.

Set Permissions, Privacy and Cost Boundaries

Learning time: about 24.1–35.1 active minutes. Choose ONE presentation: the text (about 5.1 minutes, estimated) OR the lesson video (4:04, measured). Then 20–30 minutes of practice and checking (estimated). Do not count both presentations. Setup delays and optional extra study are separate; learner timing has not yet been piloted.

Example prompt

Where you are: Module 2, lesson 4. Both assistants have a setup record. Before either builds the kit, decide which actions are allowed and which need a separate decision.

What you will do

Create a practical permission and cost boundary for your project. Recognize when a file edit, account action or usage charge is outside that boundary. Demonstrate that you can decline an unnecessary request without abandoning the task.

Teaching

“I own the project” tells the assistant who decides. It does not describe every action you authorize. Reading a synthetic CSV, changing a local calculator, emailing customers and publishing a website have different consequences. A clear brief names the actions needed now.

For the Workshop Launch Kit, the practice data is synthetic. The calculator runs locally. No names, emails, payment details or real bookings are needed. This makes the task easier to supervise and removes several unnecessary connections. Your own workplace project may involve confidential data; use only information you are authorized to share through the chosen service.

Separate written instructions from technical settings. A prompt can say “do not publish,” but the application's permission controls determine what can execute. Claude Code's current modes include manual/default, plan, acceptEdits and auto, with different behaviors. Codex exposes its own permission and approval controls. The labels are not interchangeable. Inspect the actual setting rather than assuming that “automatic” means either safe or unrestricted.

In Claude Code, /permissions opens the permission controls. In Codex CLI, /permissions lets you inspect or choose permissions, and /status shows session configuration. Use these within the relevant assistant, not at your ordinary shell prompt. A slash command belongs to its application; a Terminal command belongs to the shell.

Start with read access for inspection. Grant bounded local edits when building the practice project. Do not use permission-bypass flags to keep a lesson moving. A connector may have different controls from local shell access: a read-only folder does not necessarily prevent an authorized email connector from sending a message.

Costs also have separate layers. Subscription usage, optional API billing, third-party tool credits and hosting charges are not one allowance. A successful sign-in proves none of those budgets. Check the actual account's usage controls before a long task. Do not turn on extra usage or purchase credits just because a tool recommends it. Account limits and context capacity are different; neither is elapsed working time.

Worked example

Omar asks an assistant to improve the local workshop calculator. It requests a hosting connection. Publishing is not needed to test a local HTML file. Omar declines and asks for the local version plus a note explaining what deployment would require later. Work continues within scope.

His permission record says: “May inspect supplied inputs, edit this project and run local checks. May not send, publish, buy or access real attendee data.” A separate deployment decision remains open. Nothing in the course requires him to make it now.

Do it yourself

Write a four-row action table: action, allowed now, reason and evidence to check. Classify these requests:

  1. Read the supplied synthetic interest CSV.
  2. Edit the calculator inside the practice folder.
  3. Upload a real customer list to a newly connected enrichment service.
  4. Enable additional paid usage to avoid a limit.

Then give either assistant this complete instruction:

Example prompt

We are building a local Workshop Launch Kit from supplied fictional data. Current scope is project files and local checks only. Read the brief and propose the smallest access needed. You may not send messages, publish, buy services, enable extra billing or use real personal data. Separate file access, network access and connector actions. If a needed action is unavailable, describe the specific blocked step and a bounded alternative. Do not change settings in this task.

Compare its proposal with your table and actual application controls. Record any difference. Do not ask it to display secrets to prove authentication. Evidence can state “signed in through the official account flow” without exposing account details.

Feedback and retry

Pause here before reading.

The first two actions fit the practice scope, provided the file and folder are correct. The third introduces data and provider risks unrelated to this project. The fourth is a new financial decision. Neither follows from permission to build a local calculator.

Retry with this situation: a read-only documentation tool is unavailable, but you have a dated official-source excerpt. Choose whether to bypass controls, use the excerpt with a limitation, or claim the live source was checked. The defensible option is the dated excerpt and explicit limitation, followed by live verification when access is available.

Check your understanding

Choose an answer and write one sentence explaining why before you read the answers.

  1. A project instruction says “never send email”, while an email connector is configured with sending access. Is sending technically impossible?
    • A. No: instructions and tool permissions are different layers; inspect or restrict the connector's actual access.
    • B. Yes: the written instruction blocks it.
    • C. Yes, as long as the agent has read the instruction.

Answers

  1. A. No. Instructions and tool permissions are different layers. Keep the written boundary and inspect or restrict the connector's actual access. Do not assume one setting governs every surface.

Summary

Your boundary should let routine authorized work proceed while making new commitments visible. Keep a short record of allowed actions, data scope, cost ceiling and stop conditions. Next, you will put the project into a shared, versioned home.

Sources