Joanah Fermin

Tutorials / Disciplined AI-Assisted Development · Lesson 5 of 14

On this page

    Common Ways of Getting Better AI Results

    When an AI gets it wrong, it is tempting to blame the AI. But the AI can only act on what you told it. Making the request clear, and checking that it was understood, is the customer's responsibility, as it is at a drive-through. Lesson 03 was about being specific. This lesson is about taking that responsibility all the way: explaining well, and then checking that your message landed.

    Our goal: learn to take responsibility for being understood: explain with good context, and use five small checks that bring the AI's understanding into view before the work starts. They work in a plain chat and with an agent.

    Here is the route:

    1. A drive-through story: whose job it is to check the order.
    2. First, explain well: give the context.
    3. Five checks you run: please repeat my order, let it ask you, list the assumptions, ask for options, plan first.
    4. What these habits cannot catch, and when to skip them.

    We leave out the whole development process (lesson 06) and the documents that carry what you learn from one chat to the next (lesson 08).

    The drive-through

    You pull up to a drive-through and order: "A burger with pickles, no onions." The cashier types the opposite: onions, no pickles. You pay at the first window. At the second window, you are handed a burger with onions and no pickles. Now it costs you time, a queue of cars behind you, and a refund.

    You said it clearly, and it still went wrong. It was the customer who ended up with the wrong food, so checking is the customer's job. Before paying, say: "Please repeat my order." The cashier says: "Burger with onions, no pickles." You say: "No, that is wrong. Pickles, no onions." You fix it on the spot, before paying, and it costs nothing.

    Drive-through: without and with asking for the order to be repeated You say nothing more You: "Pickles, no onions" Cashier enters the opposite You pay Wrong food at pickup Found out at the pickup window. You ask before paying You: "Pickles, no onions" You: "Please repeat my order" Cashier: "Onions, no pickles" You: "No. Pickles, no onions" Right food Found out before paying. Same order, found out at different times.

    The mistake was the same in both stories. What changed was when it was found. With an AI the same thing is true: a misunderstanding is cheap to fix before the work, and more expensive after. And with an AI the customer's job is a little harder, because it will not tell you when it is guessing (lesson 02). If you do not check, nobody will.

    The later you find it, the more it costs What a misunderstanding costs to fix a sentence a rewrite undo thechanges Before it starts After it wrote it After an agent acted An illustration, not a measurement.

    First, explain well

    Asking for the order to be repeated catches what was misunderstood. But some mistakes start earlier, with what you left out. Ask a friend:

    Buy me water.

    Still or sparkling? Cold or lukewarm? A small bottle or a big one? Your friend has to guess, and so will an AI. People fill gaps with the usual, and an AI does the same (lessons 02 and 03). Now compare:

    Buy me 1 liter of pH 9 water, lukewarm, please.
    More context, fewer guesses Little context Buy me water. your friend guesses all four: still or sparkling? cold or lukewarm? what size? which brand? Good context Buy me 1 liter of pH 9water, lukewarm, please. stated, so no guess: 1 liter pH 9 lukewarm which brand? still a guess More context, fewer guesses. You will never remove them all.

    The second request gets the right water far more often, because every detail you state is a guess the other person does not have to make. Explaining well means giving the context a person new to the job would need:

    • What you want, and what it is for.
    • Who it is for.
    • Why: the reason behind it.
    • Limits: what it must or must not do.
    • Out of scope: what to leave alone.
    • What good looks like.

    The same goes for a software request. Compare:

    Add tags to my notes app.

    Add tags to my notes app. A tag can be up to 100 characters. A note
can have several tags. Each tag has a colour, white by default. Show
the tags under each note, and let search filter by tag. No sharing or
accounts, and keep it simple.

    Lesson 03 was about the angle you ask from. Context is about the situation: what the AI cannot know unless you say. But you cannot think of everything, and you cannot see what the AI assumed. That is what the five checks are for. You run them as the customer, to make sure your order landed.

    Habit 1: Please repeat my order

    Ask the AI to repeat your request in its own words before it does anything. Here is an example:

    The AI repeats your order, and you correct it You I want a clean database. The AI, before doing anything I plan to drop the whole database and create a new empty one. Is that right? You No. Delete all the records in every table, and do not touch anything else.

    It costs one reply. If it is wrong, you correct it before any work is done. If it is right, you have confirmed it.

    A chat would only describe that plan. An agent could carry it out, which is why this check matters most with an agent.

    Habit 2: Let it ask you

    Left alone, the AI fills any gap in your request with the typical answer (lesson 02). Invite it to ask instead:

    Ask me questions until you are sure what I want. Do not start yet.
    The AI asks before it starts You:"Add tags." The AI asks Is there a limit, up to 3 tags,or unlimited? Only the owner can tag,or other people too? Is a tag searchable? Youranswers It asks first, then it starts.

    If you want to keep it short, add "at most five questions".

    Habit 3: List the assumptions

    Every request leaves things unsaid, and the AI fills them in without telling you. Ask it to show you what it filled in:

    List what you are assuming, and what you do not know.
    What the AI assumes silently Add tags to my notes app. what you said what the AI filled in At most 3 tags per note Tags saved in the notestable, as a column Each tag up to 20 characters Tags are free text Ask: "List your assumptions."

    Some of these are choices Habit 2 asked about, such as the limit on tags. Others are technical choices you might never have thought to ask about, such as where the tags are stored. Some will be fine. Some will be wrong for you, and now you can see which. This is the gap and the fork from lesson 02: a gap is a fact it did not have, and a fork is a decision only you can make.

    Habit 4: Ask for options, and treat the first answer as a draft

    A single answer hides the alternatives. Ask for several, with their trade-offs, and choose yourself:

    Give me two or three options, with the good and bad of each.
    Options, and you choose How shouldtags work? A: free text: simple, messy B: fixed list: tidy, rigid C: both: flexible, more work You choose

    Then treat whatever you get as a draft, not the final answer. Say what to change, and ask again.

    The first answer is a draft Draft 1 Your notes Draft 2 Your notes Draft 3 Keep going until it is right.

    Habit 5: Plan first

    For anything bigger than a line or two, ask for the plan before the work:

    Show me your plan before you start. Do not change anything yet.

    In a chat, the work is writing the answer. With an agent, the work is changing your project, so a plan matters even more.

    Work straight away, or plan first Without a plan Task Work Wrong, redo With a plan Task Plan You approve Work

    A plan takes a minute to read. Undoing a wrong piece of work takes far longer, and an agent's mistakes compound (lesson 04).

    Five phrases to keep

    Five phrases 1. Repeat my order"Please repeat my order, in your words." 2. Ask me"Ask me questions until you are sure." 3. List assumptions"List what you are assuming." 4. Give options"Give me two or three options." 5. Plan first"Show me your plan before you start."

    What these habits cannot catch

    Asking it to repeat your order catches a wrong interpretation: the AI understood something other than what you meant. It does not catch a wrong fact. The AI can repeat your request perfectly and still be wrong about how a library works, because it does not flag its guesses (lesson 02).

    What asking it to repeat your order catches, and what it misses Asking it to repeat your order Catches a wrong interpretation "tags per user,when I meant per note" Misses a wrong fact "this library has afeature it does not have" For facts, you need your own knowledge, tests and a second look.

    💡 A simplification: a cashier repeating your order is a real check. An AI repeating your order is more generated text, so it can sound sure and still be off. Treat it as a good check, not a proof.

    Use the habits in proportion. A one-line rename needs none of them. A new feature deserves several, and a change that touches many files deserves all five.

    Where this goes next

    These five checks work the same in a plain chat and with an agent, which is why they come first. The course uses both. The early steps of the process happen in a chat, and from lesson 10 an agent does the building. That is where these habits matter most, and where more is needed. Later lessons add habits that only make sense with an agent.

    Habits for any chat, and habits for agents Any chat or agent this lesson Explain well Please repeat my order Ask, list, options Plan first then Only with an agent later lessons What it may do alone Teach it your project Build in small steps Check what it changed The habits here come first. The agent habits build on them.

    Recap

    • Making the request clear, and checking it was understood, is your responsibility. The AI acts on what you said, and it will not tell you when it is guessing.
    • A misunderstanding is cheap to fix before the work and expensive after. Most start with missing context, so explain well: what, who for, why, limits, out of scope, and what good looks like.
    • Then run the checks, to see how it understood you before it starts.
    • Please repeat my order: ask it to repeat your request in its own words.
    • Let it ask: ask it to question you until it is sure.
    • List assumptions: ask it to show what it filled in.
    • Options and drafts: ask for two or three options with trade-offs, and treat the first answer as a draft.
    • Plan first: ask for the plan before the work, especially with an agent.
    • Repeating the order catches a wrong interpretation, not a wrong fact, and small jobs do not need all five.
    • The checks work in a chat and with an agent. Later lessons add habits only an agent needs.

    Questions to think about

    The answers are left out on purpose. Try them in your own words first.

    1. In the drive-through story, who is responsible for checking the order is right, and why?
    2. Why does missing context make an AI, or a person, assume the wrong thing?
    3. What does asking an AI to repeat your order let you catch?
    4. Why ask for a plan before an agent changes anything?
    5. Why can repeating the order still miss a mistake?

    Next lesson: the whole process on one page. The seven steps, the two rules, and how they fit together.

    ← Agentic AI: Giving the AI HandsOverview of the Process →