- Stop polishing the sentence. Choose what goes in front of it.
- A prompt with no information in it produces the average of the internet.
- Make it interview you before it answers.
- You're briefing a new colleague, not asking a question.
- Context is where quality comes from. Role barely matters now.
- Show it, don't describe it.
- More is not better. Curation beats volume.
- Say what is wrong, not that something is wrong.
- Context controls quality — not truth, and not taste.
1. Stop polishing the sentence. Choose what goes in front of it.
Around 2023, wording genuinely mattered. Models were fragile, so "think step by step" and "you are a world-class expert" measurably helped.
Those models are gone. Anthropic's current guidance for prompting Claude says it plainly:
Prompting is converging with context engineering — less scaffolding, more curation.
What they now say to stop doing:
- Role-playing. "You are a world-class copywriter" does little, and an over-specific role can make the answer worse.
- XML tags and heavy structure. A real technique two years ago. "Less necessary with modern models."
- Long prompts. The best prompt is the one that works with the minimum necessary structure.
One thing that bullet does not mean: less thinking. You'll still write task, context and format — that's structure in what you supply, not a formula for how to phrase it. Section 5 is the whole of it.
If a course teaches you only prompt formulas, it's teaching a 2023 skill.
2. A prompt with no information in it produces the average of the internet.
"You are a world-class direct-response copywriter with 20 years of experience. Write a high-converting LinkedIn post that stops the scroll and drives engagement."
What does that tell the AI about your situation? Nothing. Not the product, the audience, what's new, or what you sound like.
So it produces the average of every LinkedIn post ever written. Which is exactly what that output feels like.
"I want to announce a new product feature on LinkedIn. Before you write anything, ask me the five questions whose answers would most change what you produce. Then wait."
One line turns a guess into a consultation. You don't have to know in advance what context matters — it tells you what it's missing.
3. Make it interview you before it answers.
- Some assistants now ask a clarifying question on a vague request, unprompted. Genuinely new and useful.
- But that is not a fixed property. Day 1, section 6 explains why — run it again and a different two may ask.
- "Claude asks, Gemini doesn't" is one run mistaken for a rule.
- One thing is stable: agent modes deliberately ask fewer questions. They're built to keep working unattended.
So don't hunt for the tool that asks. Make the ask explicit. That works everywhere, every run.
4. You're briefing a new colleague, not asking a question.
Context is everything the AI needs to do the job, that it has no way of knowing.
Picture a talented new colleague on their first morning. Widely read, fast, willing. Has never heard of your company. No access to your files. Remembers nothing from yesterday.
Hand them that "world-class copywriter" prompt and they'd stare at you. What product? Who buys it? What are we announcing?
One group is answering a question. The other is writing a brief. That's the whole difference.
5. Context is where quality comes from. Role barely matters now.
6. Show it, don't describe it.
- "Friendly but professional" means nothing. Ask five people to write it and you get five different things.
- Instead: paste two or three pieces of writing you like and say "match this voice."
- Don't describe your audience — paste three real customer messages.
- Don't describe a competitor's positioning — paste their homepage.
Give it the thing, not your description of the thing. Thirty seconds of copy-paste, and it's the fastest upgrade available to you.
7. More is not better. Curation beats volume.
- A brief is not improved by being longer.
- Dumping every document you own makes it worse — what matters gets diluted by what doesn't.
- Three well-chosen examples beat thirty.
- One paragraph on who the customer is and what they fear beats a twelve-page brand deck.
Test before pasting: would a smart new colleague actually need this to do the job? If not, leave it out.
8. Say what is wrong, not that something is wrong.
Most people write one prompt, look at the output, and judge the tool. That's judging a new hire by their first draft.
| Weak | Strong |
|---|---|
| "Make it better" | "The opening is generic. Start with the specific number." |
| "Shorter" | "Cut to 120 words. Drop paragraph three — it repeats two." |
| "More engaging" | "Reads like a press release. Rewrite as if telling one café owner over coffee." |
| "I don't like it" | "It claims something we can't back up. Use only the facts I gave you." |
Which means you have to be able to tell what's wrong. That's next week.
9. Context controls quality — not truth, and not taste.
This matters as much as everything above, and almost nobody says it.
- It won't make it truthful about things you didn't give it. A well-briefed AI invents more convincingly, not less.
- It won't make it original. Closer to your voice, further from generic — but still a blend of what it has read. The surprising angle is yours.
- It won't make it able to judge its own work. It cannot tell you whether what it wrote is any good.
Truth and taste are yours. That's what next week is about.
If you remember five things
- Context beats wording. Role-play, magic phrases and long prompts are 2023 skills.
- A prompt with no information produces the average of the internet.
- Make it interview you. "Before you answer, ask me the five questions that would most change what you produce."
- Show, don't describe — and three good examples beat thirty.
- Context controls quality, not truth and not taste. Those stay yours.
Four ways to ask for the same thing
45 minutes. Nothing to submit — one line in #daily-log at the end.
One tool, four attempts, fresh chat each time. We're changing the brief, not the tool.
The scenario
Ceylon & Co. — a small Coimbatore tea brand selling loose-leaf tea online in 100g tins.
Launching: a sample box. Six 15g tins, six teas, ₹499, free shipping. It exists because the most common objection is "₹650 is a lot for a tea I've never tasted."
Buyers: 28–45, Indian metros, already drink decent tea, curious about better. Not experts. Put off by tea snobbery.
Facts you may use: 4,200 customers. 61% of first-time buyers reorder within three months. Tea bought directly from two family estates in the Nilgiris. Founded 2021 by two sisters.
Voice: plain and warm. No "indulge in luxury". No "elevate your tea experience". They'd rather say a true specific thing than a beautiful vague one.
The job each time: an Instagram caption announcing the sample box.
Level 1 — one line (2 min)
"Write an Instagram caption for a tea brand launching a sample box."
Level 2 — the LinkedIn-template approach (5 min)
Big impressive role, lots of instruction — "world-class social media copywriter", "scroll-stopping" — and none of the scenario.
This is where most people operate. Notice whether it actually beats Level 1.
Level 3 — a real brief (10 min)
Task, context, format. Paste the whole scenario. Then paste two real Instagram captions from brands whose voice you like and say "match this voice."
Level 4 — let it interview you (10 min)
"I want to write an Instagram caption announcing a sample box for a small Indian tea brand. Before you write anything, ask me the five questions whose answers would most change what you produce. Then wait."
Answer from the scenario. Save its five questions — they're the interesting part.
Then answer these (15 min)
- Did Level 2 actually beat Level 1? Be honest.
- Level 1 vs Level 3 — name three specific things that changed.
- Level 3 vs Level 4 — which is better, and why.
- Did Level 4 ask for anything you wouldn't have thought to include? That's the whole point of the technique.
- What did not improve at any level? This is the most important question here.
Post one line in #daily-log with your answer to question 5.
Tomorrow — Day 3: Memory, tools and connectors. Today was what you put in front of the model. Tomorrow is what you bolt onto it — files, search, connectors — to make it actually do work.