From delegation to the ask itself
Last session the Meridian Retail exec team learned the mental model: AI is a brilliant, literal new hire, and output quality is set by delegation quality. This session is the practical sequel. What kinds of work can you actually hand it? What makes an instruction land? And when the answer comes back weak - as first answers often do - how do you know exactly what to fix instead of just trying again louder?
The five jobs you can give AI 8 min live
Andrew Ng's famous prompting course organises almost everything a language model does into a handful of task families: summarize, infer, transform, expand - plus the one leaders use most, create. Naming the job before you type is the single fastest upgrade to any ask, because each job wants a different brief.
LiveThe five jobs, with Meridian's actual week5 min▶
Here is each job with a request the Meridian team really made this month. Notice how naming the job makes the rest of the brief almost write itself:
| Job | You are asking it to... | Meridian example |
|---|---|---|
| Summarize | Compress long material without losing what matters | Dana: "40 complaint tickets into the top 5 themes, with counts and one quote each" |
| Infer | Read and judge: classify, extract, detect tone | The CMO: "Tag these 500 product reviews positive / negative / mixed, and pull out every mention of delivery" |
| Transform | Keep the content, change the form: rewrite, translate, reformat | Dana: "Turn this policy memo into a friendly two-paragraph email for store managers" |
| Expand | Grow a seed: bullets into a draft, notes into a memo | The CFO: "Draft a one-page board note from these six bullet points, formal tone" |
| Create | Generate options that did not exist: ideas, names, angles | The CMO: "20 distinct campaign angles for the summer sale - no two alike" |
The underused pair: most executives discover summarize and expand in week one and stop there. The quiet workhorses are infer and transform - tagging, extracting, reformatting. They are the jobs that eat the most staff hours, and the ones AI does most reliably, because the answer is checkable against the input you gave it.
Self-studyName the job before you type - the 5-second habit2 min read▶
Before your next ask, finish this sentence in your head: "I am asking it to ______ this." If the blank is summarize, lead with the material and the compression target. If it is create, lead with the quantity and the constraints. If you cannot fill the blank, the model cannot either - and you will get an answer that half-does three jobs.
- One ask, one job. "Summarize these tickets and also draft a response plan and make it funny" is three jobs. Chain them: summarize first, then expand the plan from the summary you approved.
- The job picks the brief. Summarize wants a length and an audience. Infer wants the exact labels. Transform wants the target format. Expand wants the skeleton. Create wants volume and variety rules.
The two principles 8 min live
Ng's course boils effective prompting down to two principles, and they have aged beautifully. One: write clear and specific instructions. Two: give the model room to work - context, and a reason. Everything fancier is decoration on these two.
LivePrinciple 1 · clear and specific - which is not the same as long4 min▶
The most common misreading of this principle: people hear "specific" and write longer. Length is not the point. Precision is. A 15-word ask that names the audience, the format, and the limit beats a 200-word ramble that names none of them.
| Vague (short and useless) | Specific (short and sharp) |
|---|---|
| "Make this better" | "Cut this to half its length without losing the three numbers" |
| "Write something professional" | "Write it for a board that has 90 seconds per paper" |
| "Give me some ideas" | "Give me 20 ideas, no two from the same category" |
| "Summarize this" | "Summarize this into 5 bullets a store manager can act on Monday" |
Climb the ladder one rung at a time. You rarely need rung 3 on the first try - but you should always know which rung you are standing on.
LivePrinciple 2 · give it room to work - context, and a reason4 min▶
The second principle is about generosity: give the model the material and the motive. Context you know from session 1 - paste the real thing. The motive part surprises people:
Think of any brief you have given a person. "Do it this way" gets compliance. "Do it this way, because here is what it is for" gets judgment. Models respond the same way - which should not be surprising by now, because they learned from text written by people.
Dana's why test. Dana added one sentence to her weekly ops summary ask: "This goes to the exec team who have 2 minutes; they only care about what changed and what needs a decision." The output stopped including background she already knew - without her ever listing what to exclude. The reason did the excluding for her.
Self-studyWhere the two principles come from - and why they still hold2 min read▶
These two principles anchor Andrew Ng and Isa Fulford's "ChatGPT Prompt Engineering for Developers" - one of the most-taken AI courses ever. It was written for developers, but the principles have no code in them, and Anthropic's published best practices open with the same three moves: be explicit, add context and motivation, be specific about everything that matters.
Three years of fast model progress later, plenty of 2023 tricks have died (session 1 covered the graveyard). These two principles did not - because they are not model tricks. They are delegation, formalised. As models improve, sloppy asks get better answers too, but the gap between a vague ask and a sharp one stays wide. It is the one gap the vendors cannot close for you.
Read the miss 6 min live
Session 1 gave you the anatomy of a strong ask: goal, context, constraints, format, example. Here is the payoff. A weak answer is not a verdict on the tool - it is a symptom, and the symptom tells you exactly which ingredient your brief was missing. Diagnose, add one thing, go again.
LiveThe diagnosis table - from symptom to one-line fix6 min▶
| Symptom - what came back | Missing ingredient | The fix |
|---|---|---|
| Right topic, but generic - could be about any company | Context | Paste the real document, numbers, or situation. "Our" means nothing to a day-one hire. |
| Beautiful answer to a question you did not ask | Goal | Say what done looks like: the decision it supports, the audience it serves. |
| Three pages when you needed three lines; or hedged, jargon-heavy, off-limits topics included | Constraints | State length, tone, and the rules: what to include, what never to mention. |
| Content is fine but the shape is unusable - prose when you needed a table | Format | Name the shape: "a table with these 3 columns", "an email under 150 words". |
| Right shape, wrong voice - bland, or nothing like how you write | Example | Paste one gold sample and say "match this". One example beats ten adjectives. |
The discipline: add only the missing ingredient. Rewriting the whole prompt from scratch each round teaches you nothing. Fixing one named gap per round teaches you the craft in a week.
Self-studyWhy "try again but better" fails - and diagnosis works2 min read▶
When an answer disappoints, the untrained reflex is to re-ask with more emphasis: "No, really summarize it WELL this time." That adds emotion, not information - and the model already tried its best with what it had. The trained reflex is a 10-second diagnosis: which row of the table is this? Then one surgical addition.
This is the iterative loop from session 1, now with a steering wheel. First answer, read the miss, add the ingredient, second answer. Two rounds of diagnosis routinely beat five rounds of frustration - and unlike frustration, the skill compounds: after a month you pre-empt the miss and write the ingredient in the first ask.
The makeover: the CMO's campaign brief ★ 15 min · work it live
In session 1 you heard how the Meridian CMO pasted a one-liner, got bland copy three times, and went back to the agency. Today she gets her makeover. Watch the anatomy assemble - then run the same upgrade on one of the three real asks you brought.
Run the one-liner first. Send Prompt A as-is. Keep the bland result open in a tab - it is your before photo, and it is what "the AI is mediocre" actually looks like from the other side.
Name the job. This is create plus a little transform: new options, in a defined voice. That tells you the brief needs volume rules, a voice sample, and format constraints - the taxonomy just wrote your checklist.
Assemble the anatomy. Audience, the why, the voice sample, format, constraints. Notice the sample caption does more work than any adjective - "warm, playful, family-focused" describes it; the sample IS it.
Compare out loud and name the winner. Which ingredient moved the output most? For voice-driven work it is almost always the example. For analytical work it is almost always context. Knowing your pattern makes the next brief faster.
Now yours. Take the best of your three real asks and climb it to rung 3. Golden-rule it with your neighbour: if they can execute your brief cold, so can the model.
Three things before session 3 ◐ 20 min total
- Name the job on five real asks this week - say "this is summarize" or "this is infer" out loud before typing. Notice how the brief writes itself once the job has a name.
- Build a voice sample pack: collect 3 gold examples of writing you want AI to match - your best email, your best memo, a caption you loved. Pasting one of these is your highest-leverage single move.
- Diagnose one miss with the table. Take one disappointing answer, find its row, add only the missing ingredient, and re-run. Bring the before and after - session 3 opens with the best one.
What this session covers
Leader-level coverage of the sources below - working ideas in plain language, honestly flagged where the depth lives in the builder track. Official certificates and videos stay with the vendors; links are in the course map.
Three questions before you go 🎯 ◐ 90 seconds
1 · Dana wants 500 store-visit reports turned into a one-page brief for the exec team. Which job is she delegating?
Long in, short out is summarize - so the brief needs a compression target ("one page") and an audience ("exec team"). Naming the job tells you what the brief must contain.
2 · "Write clear and specific instructions" means your prompts should be...
Specific is not long. "Cut this to half without losing the three numbers" is nine words and rung-3 sharp. Length without precision is just a longer vague prompt.
3 · The AI returns a polished answer about retail in general - nothing about your company. Which ingredient was missing?
Generic output is the signature symptom of missing context. The model filled the gap with what it knows about retail at large. Paste the real material and the answer becomes yours.