{{ c.h }}

{{ c.p }}

Practice this

{{ dt.tick }}

{{ dt.label }}

{{ drill.name }}

{{ drill.brief }}

How to do this drill

1 · Every card has to end up in a column. There is no "leave it here" option — the holding area at the top must be empty before you can check.

2 · Drag a card into a column, or tap the card and then tap the column. Both work.

3 · Put a card back by tapping it in its column, any time before you check.

4 · Then press Check my sort. You will get the reasoning on every card, right or wrong, and you can reset and retry as often as you like.

{{ clSortStatus }}

{{ it.text }}

All sorted. Check your answers below.

{{ b.label }}

{{ b.hint }}

{{ p.text }}

{{ p.why }}

Check my sort Reset

{{ clResult }}

{{ clStageNote }}

{{ rpHead }}

{{ rpIntro }}

The situation

{{ rpScenario }}

{{ rpTo }}

{{ f.label }}

{{ f.hint }}

{{ f.verdict }}

Check my report

{{ rpResult }}

A report a responder can work from

{{ rpModel }}

{{ rpModelNote }}

How to do this drill

1 · Three colleagues each wrote a complete prompt for the same task. Read all three before judging any.

2 · For each prompt, mark your verdict: Weak, Mixed, or Strong. You are judging, not writing.

3 · This is the skill that comes before building — recognizing what a good prompt looks like and where a weak one fails.

4 · Press Check my judgment to see each prompt’s true strength, the draft it would produce, and why the weak ones fail.

{{ critStatus }}

{{ p.label }}

{{ p.verdictMark }}

{{ p.prompt }}

{{ o.label }}

{{ p.correctLabel }}

{{ p.weakCoaching }}

{{ p.responseLabel }}

{{ p.response }}

Check my judgment Reset

{{ critResult }}

{{ pcTitle }}

{{ pcDisclosure }}

What is happening in this drill

There is one AI tool in this drill and one prompt. You are not comparing ChatGPT against Claude against Copilot — you are comparing badly written instructions against well written ones, given to the same tool.

A prompt is just the message you type. Here it has been broken into {{ pcBlockCount }} blocks — {{ pcBlockList }} — because those are what separate a prompt that works from one you have to rewrite. Each block has a short explanation above its options.

Under each block are three versions of that same block: one weak, one halfway, one strong. Three versions, not three tools. Choosing one writes it into the prompt you can see assembling further down the page — that grey box is literally what would get sent.

When all {{ pcBlockCount }} blocks are chosen, press See what this prompt returns. You get back the answer that prompt produces, plus a one-line verdict on every block you picked, so you can see which specific choice caused which specific weakness in the answer.

Then change one block and run it again. Watching the answer get worse when you weaken a single block is the whole point of the exercise, and it takes ten seconds.

How to do this drill

1 · Nothing here needs your own information. Every option is already written for you — your job is to choose between versions, not to type.

2 · Pick one option in each of the {{ pcBlockCount }} blocks. They differ in quality, not in subject, and only one of the three is strong.

3 · Watch Your assembled prompt build underneath as you choose. That is the single message being sent to one AI tool.

4 · Press See what this prompt returns. You get the answer that prompt produces, plus a verdict line per block: ✓ strong, ! halfway, ✕ weak.

5 · Change any one choice and run it again. Seeing a weak block spoil the answer is the exercise — there is no penalty for running it ten times.

{{ pcStatus }}

{{ f.label }}

{{ f.hint }}

{{ f.what }}

{{ o.text }}

How to do this drill

1 · Write your own text in every box. The faint grey text is an example to show you the shape — it is not an answer, and it will not be submitted.

2 · Use your own work, because the point is to leave with something you will actually reuse. Keep to the data rules: use [SLOTS] instead of real customer names, pricing or figures.

3 · The button below stays greyed out until every box has something in it.

4 · Stuck on a box? Show a strong example opens a full worked version you can borrow the structure from.

{{ pbFillStatus }}

{{ f.label }}

{{ f.hint }}

{{ f.exampleNote }}

Your assembled prompt

{{ pbAssembled }}

{{ pbRunLabel }} {{ pbExemplarLabel }} {{ pbSaveLabel }} Reset Manual override — mark as passed

{{ pcOverrideNote }}

{{ pcOverriddenNote }}

{{ r.mark }}

{{ r.text }}

{{ pbOutputLabel }}

{{ pbOutput }}

A strong version of this prompt — same tool, better instructions

This is one worked prompt written out in full, for the same task and the same AI tool. Read it against the prompt you assembled above and look for what yours leaves to guesswork.

{{ drill.exemplar }}

The three tiers

{{ t.label }}

{{ t.line }}

{{ t.examples }}

How to do this drill

1 · Read each task and choose the tier you would actually reach for. There is no penalty and you can change your answer.

2 · The moment you choose, all three answers open up — what the low, medium and high tier each returned for that task, side by side.

3 · Read the losing answers, not just the right one. Seeing the fast tier invent a figure, and the top tier over-produce, is the lesson.

4 · The verdict under each task explains which tier the task needed and why. Answer all three tasks to complete the drill.

{{ mpStatus }}

Illustrative, not live

{{ mpNote }}

{{ t.idx }}

{{ t.complexity }}

{{ t.title }}

{{ t.text }}

{{ o.label }}

{{ o.line }}

{{ t.verdictLabel }}

{{ t.verdict }}

{{ r.label }}

{{ r.flag }}

{{ r.examples }}

{{ r.text }}

{{ r.note }}

Where the tier names come from

{{ s.t }}

Reset

How to do this drill

1 · Read the agent configuration and the task, then press Run the agent. Do not study the document first — the point is to meet the output the way you would on a normal Tuesday.

2 · The output is wrong in three ways. Work out how, then find the sentence in the document that caused it. Only full sentences are clickable, and the wrong guesses are all real findings.

3 · Once you have found it, choose how you contain it. Three of the four options are things people actually do, and they all make it worse.

{{ ilStage }}

The agent, as configured

{{ c.t }}

Your instruction this morning

{{ ilTask }}

{{ ilDocTitle }}

{{ s.text }}

Run the agent

Nothing is sent anywhere. The result below is authored, so the lab works with no model attached.

{{ ilOutHead }}

{{ ilOutText }}

Three things wrong with that

{{ t.t }}

Your turn · click the sentence that did this

The questionnaire is below and the underlined sentences are now clickable. Pick the one that redirected the agent. A wrong pick is explained, not punished — every decoy is a real finding you should raise with the vendor.

{{ ilDocTitle }}

{{ s.text }}

{{ ilTries }}

{{ nt.head }}

{{ nt.why }}

Contain it

You have the run, the output and the sentence. One of these four is containment. The other three are what gets chosen under time pressure.

{{ f.label }}

{{ f.line }}

{{ f.why }}

{{ ilResult }}

Reset the lab

How to do this drill

1 · Read the passage below. Only the underlined parts can be clicked — the rest is there for context.

2 · Click every underlined claim you would not send out without checking it first. Leave the ones you would be comfortable standing behind.

3 · Flagging nothing and flagging everything both score badly. The skill is telling them apart.

4 · Press Check my flags for the reasoning on every underlined claim, including the ones you left alone.

{{ s.text }}

Check my flags Reset

{{ halResult }}

{{ n.head }}

{{ n.why }}

How to do this drill

1 · Three situations below. Pick the tool you would reach for first in each — one choice per situation.

2 · All three tools are approved, so this is not a safety question. Judge it on where the data already lives and how easily you could check the output.

3 · The reasoning appears as soon as you answer, and you can change your answer afterwards.

{{ c.s }}

{{ o.label }}

{{ c.head }}

{{ c.why }}

{{ toolResult }}

Reset

How to do this drill

1 · The claim below is already filled in for you. Leave it as it is — you do not need to supply anything of your own.

2 · Run Pass A, then run Pass B. Same claim, two different instructions — two separate conversations, not two different products. In this build no model is attached, so both answers are worked examples written for the course.

3 · Read them side by side. The point of the drill is the difference between them, not either answer on its own.

4 · Write what you noticed in the log, in your own words, then press Log it and finish. A sentence or two is enough.

The claim to check

Supplied for you. Editable, but you do not need to change it.

Two passes over the same claim. Which product you use matters less than the fact that the second pass is independent of the first — fresh conversation, different tool where you have one, and never told what the first pass concluded.

Pass A — plain ask

First conversation. Ask it straight, the way you normally would.

{{ vLabelA }}

{{ vOutA }}

Pass B — adversarial second opinion

Fresh conversation, told to attack the claim. At work, run this one in a different vendor’s tool.

{{ vLabelB }}

{{ vOutB }}

Your discrepancy log

Where did the two passes disagree, and what would you check with a human before this leaves the building?

Log it and finish

{{ vResult }}

Reset

{{ drill.guidance }}

Goal

{{ drill.goal }}

How to do this drill

1 · Every step below has to be sorted. Nothing may be left in Available steps.

2 · Steps the agent should run go in Your chain, in the order it should run them. Use the arrows to reorder.

3 · Steps the agent should not run go in Would not use. Some steps here are deliberately unsafe, and putting them in Would not use is the correct answer — not a way of skipping them.

4 · Then press Review my design. You can move any step back out and try again as often as you like.

{{ agSortStatus }}

Available steps · {{ agAvailCount }}

{{ p.label }}

Add to chain Would not use

Every step is sorted. Press Review my design.

Your chain · runs in order

{{ c.pos }}

{{ c.label }}

↑ ↓ ✕

Nothing here yet. Add the steps the agent should actually run.

Would not use · agent never runs these

{{ r.label }}

✕

Nothing excluded yet. Steps you would refuse to give the agent belong here.

Review my design {{ agRunLabel }} Clear

{{ f.mark }}

{{ f.text }}

Dry run — plan only, nothing executed

{{ agRunOut }}

{{ q.n }}

{{ q.q }}

{{ o.letter }}

{{ o.label }}

{{ q.why }}

Submit Retry