NewAI Content Creation is now live in Early Access
Turning Point Training Instituteby Training Center
Skip to content
0%
Module 1 · Own the projectLesson 1 of 2
Video lesson

L01 · From a checked workflow to a project you own

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.

From a checked workflow to a project you own

Learning time: about 18.7–30.4 active minutes. Choose ONE presentation: the text (about 5.4 minutes, estimated) OR the lesson video (3:43, measured). Then 15–25 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 1, lesson 1 of 2. You are bringing the checking habits from Practical AI Workflows into a project that agents can help build.

What you will do

Choose a bounded project, explain the roles of its human owner and agents, and identify evidence that distinguishes a working result from a confident completion message. This begins outcome O1.

Teaching

In the first course, you turned a manager's request into a checked recommendation. You retained actual AI output, inspected sources and numbers, tested changed inputs and prepared a usable handover. Keep those habits. Here your agents can also change files and run commands. Their actions therefore need a clear place to happen, a limited assignment and a way to inspect the result.

Our fictional project is the Workshop Launch Kit. A training coordinator needs to explore the budget for a small workshop. The finished learning project will include a brief, a calculation with synthetic inputs, a local calculator and a handover. It will help the coordinator inspect scenarios. It will not book a venue, charge a customer, publish a website or establish demand for a real course.

The budget uses USD 180 fixed cost, USD 8 variable cost per attendee, USD 30 ticket price and capacity 12. Fixed cost applies once. Variable cost grows with attendance. Interest data, when supplied, is fictional interest rather than paid bookings. You must keep that distinction in the interface and recommendation.

An agent can write a calculator that opens successfully and still applies the wrong formula. It can produce a correct calculation while hiding an important assumption. Your responsibility is to judge whether the artifact serves the agreed purpose. You do not need to write every line of code yourself, but you must be able to explain what the project accepts, what it returns and how you checked it.

Claude Code and Codex are the two agents you will use in this course. Their exact interface and access depend on your environment. You will verify those details in Module 2. GitHub can hold shared versioned files. It does not make the agents talk to each other automatically. A handoff must name the version, completed work, checks and next assignment. MCP connections can expose tools to an agent; an installed connection does not show that the tool was successfully used.

For now, use a document. Separate human decisions, agent assignments and observable results. The human chooses the project's purpose and acceptable use. The builder produces a bounded change. The reviewer inspects that change against evidence. The owner decides the next action. Later you will practise these roles without letting two agents edit the same files at once.

Worked example

A coordinator asks, “Make something that helps me decide whether this workshop is affordable.” A useful response is a local calculator whose inputs and assumptions are visible. A test can enter eight attendees and check the resulting budget against separate arithmetic.

The builder reports, “Finished and tested.” Before accepting that statement, ask what was opened, which values were entered, what result appeared and where the result was recorded. If the browser was unavailable, the record must say that. A source-code inspection and a browser interaction are different checks.

Do it yourself

Write PROJECT-INTENT.md in your own learning workspace, or a plain document until setup is complete. Include the intended user, one decision the kit supports, the artifact, the permitted fictional inputs, three actions outside scope and two observations that would demonstrate a useful result.

Then use this complete request in an authorized assistant if available:

text
Review this learning-project intent as a skeptical colleague. The project is a fictional Workshop Launch Kit for a training coordinator. It uses USD 180 fixed cost, USD 8 variable cost per attendee, USD 30 ticket price and capacity 12. It will create local files only. No booking, payment, sending, deployment or real personal data is authorized. Interest is fictional and is not bookings or evidence of market demand.
Read my intent below and identify ambiguous deliverables, missing observable checks and any action that exceeds scope. Do not build, modify files or claim tests ran. Return findings and the smallest suggested clarification.
MY INTENT:
[Paste your complete intent document here.]

Replace the input slot before submitting. Retain the response and your decision on each suggestion. If AI is unavailable, critique it yourself and mark the interaction pending.

Feedback and retry

Pause before reading. A sound intent makes the output and its use clear, excludes external commitments and distinguishes fictional inputs from demand evidence. “Build the best system” cannot be checked. Replace it with a behavior someone can observe. If your intent permits the agent to publish or email, remove that permission for this learning project and explain the correction.

Retry with one new condition: the coordinator cannot access the internet during review. State which artifact should still work locally and which source checks must be recorded beforehand. Do not promise an offline feature you have not tested.

Check your understanding

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

  1. Does access to the same GitHub repository prove that Claude and Codex exchanged an assignment?
    • A. No: inspect the explicit handoff and the exact shared version.
    • B. Yes: both agents can see the same files, so the assignment was shared.
    • C. Yes, if both agents say they read the repository.
  2. Can a successful file write prove that a calculator's result is correct?
    • A. Yes: if the file saved, the code works.
    • B. Only if the agent reports no errors.
    • C. No: compare actual behaviour with independently expected results.

Answers

  1. A. No. Inspect the explicit handoff and the exact shared version.
  2. C. No. Compare actual behaviour with independently expected results.

Summary

Keep the project small enough to inspect. Your first deliverable is a clear intent with evidence requirements, not permission for an agent to act everywhere.

Sources

GitHub, GitHub flow, reviewed 8 November 2026, supports versioned collaboration. MCP, architecture overview, reviewed on the same date, distinguishes exposed capabilities and execution. The Workshop Launch Kit and role division are original course teaching examples. First-course prerequisite mapping in the course production record (not a learner download).