Principles of AI · Unit 2: Working With AI · Project 2.2.4

Tool Decision Memo

🛠 Project · Individual + partner · ~3 days
Break a real task into steps, decide which ones AI gets and which stay yours, test the shakiest AI step, and write the memo that defends your split.
Back to Project 2.2.4

🎯The Challenge

read this first

"Use AI for this" isn't a plan. A real task has steps, and AI fits some of them and not others. Pick one real task you will actually do soon, break it into 5 to 8 steps, and decide for each step who does it: you, AI as a drafter, AI with your review, or AI on its own with a log. Then test the AI step you are least sure about and let the result change your mind. The product is a Tool Decision Memo that defends your split with evidence.

Rules of the project
  • Rule first (2.1.3). Check the rule that applies to you. If it says no, pick a different task. If it doesn't cover your use, ask, and write down the answer.
  • A real task that isn't graded. Pick something you will actually do: an event, a study plan, a proposal, a presentation, a budget. Don't use a graded assignment unless your teacher says you may.
  • Private information stays out of AI. That means names, addresses, grades and other people's details, and anything you're not sure about. Those steps stay with you.
  • Health or safety steps. If your task touches allergies, medicine or safety, AI may draft but a person verifies against an authoritative source. The page won't let AI act on those steps.
  • No real agent on your real accounts. You may recommend an agent-style tool in your memo, but you may only test one in the Sandbox (2.2.3).
  • Decide before you test. Write your criteria and choose your lane first, so the test can change your mind.
5 of 18 checks green · your work saves in this browser as you type

🧰Decision Toolkit

everything new in this project

You already have most of this: checkable criteria and logs (2.2.1), claims you can defend (2.1.3, 2.2.2), and the ladder from read to hard to undo (2.2.3). These tools put them together. Open each one when you reach that step.

The Split Test: five questions per step
  • Can I check the result quickly? If you can't check it, you can't review it.
  • If it's wrong, can it be undone? This is the ladder from 2.2.3.
  • Does it need my judgment or values? Taste, fairness, knowing the people involved.
  • Does it involve private information? Yours or anyone else's.
  • Is it routine and repetitive? A job a script could almost do.

"Not sure" counts as the careful answer for judgment and private information (yes), and for checking and undo (no). Go find out, then answer again.

Check in this orderLane
Private information? Judgment needed and it can't be undone? Can't check it and can't undo it?Me only
Needs judgment (but can be undone)? Or can't check it but can undo it?AI drafts, I decide
Checkable, no judgment, but can't be undone? Or checkable, undoable, not routine?AI does it, I review first
Checkable, undoable, no judgment, no private information, and routineAI does it, logged

The Split Test is a starting point. You may choose a lane with less AI than it suggests, with a reason. You may choose more only if your own answers allow it, and you must explain it.

The four lanes and the settings from 2.2.3
LaneWhat it meansSetting in the Sandbox
Me onlyAI isn't involved.AI is Off for this step
AI drafts, I decideAI suggests or drafts. You edit, decide and do anything that acts.AI may draft; I decide and I do any acting
AI does it, I review firstAI does the step, but it doesn't count until you've reviewed it.Ask first
AI does it, loggedAI does the step and keeps a log, with no approval needed.Auto, with a log
Tool types: what each is good for, and what to watch
Tool typeGood forWatch out for
ChatbotDrafting, rewording, brainstorming, explaining an idea in plain words.Facts, numbers and sources can be invented. Never put private information in. Check what matters (2.1.1–2.1.3).
Search toolFinding current sources and checking a claim.A summary can be wrong or out of date. Open the sources, and check the date (2.2.2).
Spreadsheet or calculatorExact arithmetic, sorting, totals, schedules. The same input gives the same answer every time.A wrong formula is still wrong. Test it on one example you can do by hand.
Image or media generatorPictures, posters and drafts you will look at and judge.Check the tool's terms for sharing. Describe what you want in your own words, no real people, and no living artist's name (2.2.2).
Agent-style tool (it acts)Steps that involve acting on several tools in order.Hard to undo. This course only tests agents in the Sandbox (2.2.3), so if you choose this, say you have not tested it.

The best tool for a step is sometimes not AI at all. A spreadsheet adds numbers the same way every time. A person who knows the people makes the call.

A mini-test (like the stations in 2.2.1)
  • Pick the AI step you are least sure about, not the easiest.
  • Write two yes/no criteria first, so a stranger could check them.
  • Round 1 is your prompt as you'd really write it. Each round after changes one thing. Stop at both criteria met, or after three rounds.
  • Count all the time, including checking and fixing (2.1.2), and compare it with your estimate of doing the step by hand.
  • A test with two rounds is a hint, not proof. Say so in your limits.
An invented example, worked by the same rules

Plan the robotics club's bake sale (invented). Answers are in the order check · undo · judgment · private · routine.

#StepAnswersThe Split Test suggests
1Choose the goal and the price listno · yes · yes · no · noAI drafts, I decide
2Reformat the shopping list into a table sorted by store aisleyes · yes · no · no · yesAI does it, logged
3Assign volunteer shifts using classmates' namesyes · yes · yes · yes · noMe only
4Post the announcement on the club's public pageyes · no · no · no · noAI does it, I review first
5Estimate how many people will comeno · yes · no · no · noAI drafts, I decide
The memo: what goes in it
  • Recommendation first. Which steps use AI, which don't, and why, in a few sentences.
  • The split with the Split Test answers for each step, so anyone can see how you got there.
  • Evidence from your test, with the criteria you wrote first and your time.
  • Risks and guardrails for steps where AI acts, and a fairness check from 2.2.2.
  • The strongest objection to your split, and your reply. Then your limits and your AI-use statement, including your decisions.

1️⃣The Brief

Day 1

Ideas: plan a club event or fundraiser, a study plan for finals week, a presentation, a team schedule, a proposal to a teacher, a budget for a trip. Pick something with 5 to 8 steps that you will really do.

Rule check (2.1.3)

2️⃣Steps and the Split Test

Day 1

Break the task into 5 to 8 steps, in order. Do the steps yourself first, without AI, so you know what each one really involves. Answer the five questions. The page suggests a lane. You choose.

1.
2.
3.
4.
5.
5 of 8 rows
Your split so far (computed)

Describe your steps to see your split.

3️⃣Test the Shakiest AI Step

Day 2

Test one step (or two) where AI is involved. Write your two criteria first. Round 1 is your prompt as you'd really write it. Every round after that changes one thing. Record whether each criterion was met and how many minutes the round took, including checking and fixing. Use an approved chatbot or tool with no private information, or do the step with a spreadsheet.

You need a step in lane "AI drafts," "review first," or "logged" to test. Choose lanes above first.

Test 1
R1
R2
R3
Test 2 (optional)
R1
R2
R3

4️⃣Fairness and Tools

Day 2

From 2.2.2: allowed and fair are different questions. Answer for your whole task.

5️⃣Peer Challenge

Day 3

Swap with a partner. They read your split, then pick the step they would put in a different lane and say why. You reply, and say whether it changed your split. A good challenge points at a Split Test answer, not at a feeling.

6️⃣The Memo

Day 3

The split, evidence, fairness, challenge and statement fill themselves in from your work. You write the rest.

AI-use statement (2.1.3, with your decisions from 2.2.1)
Checks (computed from your work)
The brief is complete. Name the task, who it is for, two checkable ways to tell it is done, and the worst realistic mistake (5 words or more).
A rule clears the task. Pick what your rule says and copy the exact words, or what the answer was when you asked.
5 to 8 steps, each described. You have 0 described steps.
All five Split Test answers for every step. Answer all five questions for each step.
A lane is chosen for every step. Choose a lane for each step.
No step has a lane the Split Test rules out. Every chosen lane is allowed by your answers.
Every change from the suggested lane has a reason. Where you chose differently from the Split Test, you explained why.
Each AI step has a tool type. Every step that uses AI has a tool type.
Steps where AI acts have a risk and a guardrail. Each step where AI acts has a risk and a guardrail.
At least one valid test, and every test you started is valid. Test a step in lane D, R or A: both criteria, your by-hand minutes, at least two rounds in order, and minutes for each round.
Each test ends with a lane, and keeping a higher lane than the test supports is explained. Every test ends with a lane you can defend.
The fairness check is answered. Answer all four fairness questions.
Tools are named and approved. Name the tools you would use, and confirm your teacher or district allows them and their terms allow students your age.
A recommendation. Say what you recommend (at least 20 words): which steps use AI, which don't, and why.
The strongest objection and a reply. State the strongest objection to your split and reply to it, 8 words each.
Limits. What can't your test show? At least 15 words. Include the tool, how few rounds you ran, and anything you didn't test.
AI-use statement. Fill in all six parts, including your decisions.
A peer challenge and your reply. A partner picks one step and challenges it (5 words or more). You reply (5 words or more).
Your memo · not finished yet
TOOL DECISION MEMO
Task: …
Done looks like: (1) … (2) …
Worst realistic mistake: …
Rule check: … — …

RECOMMENDATION
…

THE SPLIT (0 steps: 0 involve AI, 0 stay with me · 0 draft, 0 review-first, 0 logged; 0 differ from the Split Test suggestion)

EVIDENCE
(no test yet)

FAIRNESS (2.2.2)
Uses other people's work: ? · names or shows people: ? · could replace someone's work: ? · will be shared beyond class: ?

TOOLS: … · allowed by my teacher or district and its terms: not confirmed

STRONGEST OBJECTION
…
REPLY
…

LIMITS
…

PEER CHALLENGE
(none yet)

AI-USE STATEMENT
Tool: … · Used for: … · Not used for (I did this myself): … · What I checked, and how: … · Policy that applies: … · My decisions: …

📐What Strong Work Looks Like

PartMeets the projectStrong
TaskA real, small task with 5 to 8 steps, cleared by the rule, with two checkable ways to tell it's done.You can say what the worst realistic mistake is and who it would affect.
SplitEvery step has all five answers and a lane. Overrides have reasons.Your reasons point to the answers, and you can name a step where the Split Test surprised you.
TestOne valid test with criteria written first and honest time, including checking.You tested the step you were least sure about, and the result changed a lane or gave a clear reason it didn't.
GuardrailsSteps where AI acts have a risk and a guardrail.Your guardrails fail in different ways, and you say how.
FairnessAll four questions answered, with a plan for any yes.You separate what's allowed from what's fair.
ChallengeA partner challenged a step and you replied.Your reply cites evidence, or you changed your split and say why.
HonestyLimits are specific.You claim nothing about AI in general, only about your task, your tool and your test.

🛠 Project 2.2.4 · Unit 2 of 3 · 11 Weeks