The 6 Claude Prompts That Put You Ahead of 99% of AI Users

Claude has thousands of capabilities buried inside one empty text box. Most people never use more than one of them: type a question, get an answer, close the tab.

That's it. That's how the vast majority of people use the most capable tools ever put in front of office workers. They treat a reasoning engine like a search bar, get a mediocre paragraph back, and conclude "AI is overrated."

Here's what I've learned after two years of using Claude, ChatGPT, and Gemini every single working day: the gap between average users and power users is not technical knowledge. It's six sentences.

Not six frameworks. Not six paid courses. Six short prompts you can memorize in an afternoon, that work in every major AI tool, and that change the shape of the answers you get — from generic filler into work you'd actually put your name on.

In this post I'll give you all six. For each one you'll get:

• The exact prompt, ready to copy and paste
• Why it works (in plain English, no jargon)
• Three real workplace scenarios — email, meetings, planning, research
• How to chain it into a repeatable weekly workflow, because a prompt you use once is a party trick and a prompt you use every Monday is a system
• Variations for when the basic version isn't quite right

One note before we start. Everything below works in Claude, ChatGPT, and Gemini. I write "Claude" because that's my daily tool, but if your company gives you Copilot or Gemini, nothing changes. These prompts work on the model, not the app.

Let's go.

The six prompts at a glance

#	Prompt	What it fixes	Best moment to use it
1	"Explain it like I'm 10."	Answers you don't actually understand	Before a meeting where you have to sound informed
2	"Ask me clarifying questions before answering."	Generic answers to vague requests	Any task worth more than 10 minutes
3	"Show me 3 examples first."	Getting one mediocre draft and settling	Anything with taste involved: writing, naming, design
4	"Save this as a reusable skill / instruction."	Repeating yourself every session	The third time you type the same request
5	"What are my blind spots here?"	Confident plans with hidden holes	Before you hit send, ship, or commit
6	"Don't stop until [definition of done]."	Half-finished, 80%-there output	Long tasks the AI tends to abandon early

Print that table. Stick it next to your monitor. Now let's take each one apart.

Prompt 1: "Explain it like I'm 10"

Explain it like I'm 10.

That's the whole prompt. Append it to anything.

Explain what a vector database is like I'm 10.

Here's the pricing section from our new vendor contract. Explain it like I'm 10,
then tell me the one clause a careful adult would worry about.

Why it works

AI models mirror the register of the question. Ask a technical question in technical language and you get a technical answer — which is useless if you didn't fully understand the technical language in the first place. Most people nod along to an answer they half-understand, and half-understanding is where bad decisions are born.

"Explain it like I'm 10" forces the model to do the thing genuinely smart people do: strip away vocabulary and keep the mechanism. It can't hide behind terms like "leverage synergies across the data layer." It has to say "imagine a library where books are shelved by what they mean, not by their titles."

There's a second, less obvious benefit: it's a comprehension test for you. If the simple explanation surprises you, you didn't understand the complicated one. I catch myself with this weekly.

Three workplace scenarios

Scenario 1 — The pre-meeting rescue (email/calendar). You have a 2 p.m. call about "the SSO migration" and you're not entirely sure what SSO is beyond "the login thing." Ten minutes before the call:

Explain SSO like I'm 10. Then give me the 5 terms most likely to come up
in a meeting about migrating a company to SSO, each in one plain sentence.
Then list 3 smart questions a non-technical manager could ask.

You walk in able to follow the conversation and ask one question that lands. That's the entire game of cross-functional meetings.

Scenario 2 — The dense document (research). Legal sends over a 14-page data processing agreement. Paste it in:

Explain this agreement like I'm 10, section by section. Flag any section
where the simple version sounds bad for us.

Before: you skim, feel vaguely anxious, sign anyway. After: "Section 7 basically says: if they lose your data, the most they'll ever pay you is one month's fee." Now you know exactly what to push back on.

Scenario 3 — Explaining your own work upward (planning). This one is the sleeper hit. Use the prompt in reverse — on your material:

Here's my project update for the executive team. Rewrite it like the readers
are smart 10-year-olds with 30 seconds of attention. Keep every number.

Executives are not 10-year-olds, but they are reading your update on a phone between meetings. The simplified version is almost always the better version.

Chain it into a weekly workflow

Here's my repeatable "understand anything" chain — I run it every time a new topic lands on my desk:

1. Explain [topic] like I'm 10.
2. Now explain it like I'm a smart college student. What did the simple version leave out?
3. What do people most often get wrong about this?
4. Give me a 3-sentence version I could say out loud in a meeting without embarrassing myself.

Fifteen minutes, and you've gone from zero to conversational. Save those four lines in a note called "Learn Anything Ladder" and reuse them weekly. (If you want the deeper version of this method, it's the backbone of my complete prompt engineering guide.)

Variations

Variation	When to use it
"Explain it like I'm new to the team"	You want plain language but adult context
"Explain it with an analogy from cooking / sports / driving"	Abstract topics that need something concrete
"Explain it like I'm 10, then like I'm an expert — show both"	You want to see exactly what the jargon is hiding
"Explain it to my boss in 4 sentences"	Turning your understanding into upward communication

Prompt 2: "Ask me clarifying questions before answering"

Before you answer, ask me clarifying questions — as many as you need
to do this well. Wait for my answers.

If I could teach the average professional exactly one prompt, this would be it.

Why it works

Think about how a great consultant, lawyer, or contractor behaves. You say "I want to redo my kitchen," and they don't start swinging a hammer. They ask: What's the budget? Who cooks? Gas or electric? Are you selling the house or staying twenty years?

AI models, by default, do the opposite. They're trained to be helpful immediately, so they take your vague request, silently guess at every missing detail, and produce an answer to a question you didn't quite ask. The guesses are invisible — that's the dangerous part. You get 600 confident words built on assumptions you never got to veto.

This prompt flips the default. Instead of guessing, the model surfaces its assumptions as questions. You answer in 60 seconds of shorthand, and the quality of the final output jumps more than any other single change you can make. I'm not exaggerating: vague prompt plus clarifying questions beats a detailed prompt written cold, because the model asks about things you'd never think to specify.

Three workplace scenarios

Scenario 1 — The high-stakes email. You need to tell a client their project is slipping by three weeks.

Before (the usual way):

Write an email telling the client the project will be 3 weeks late.

You get a corporate apology template. Generic, groveling, forgettable.

After:

I need to email a client that their project is slipping 3 weeks.
Ask me clarifying questions before drafting anything.

Claude comes back with: Whose fault is the delay? Have they been warned before? What's the relationship like — formal or friendly? Is there anything you can offer to soften it? What reaction are you worried about? You answer in bullet-point shorthand — "partly their fault (late feedback), warm relationship, can offer a free month of support, worried they escalate to their CFO" — and the draft that comes back sounds like it was written by someone who actually works on the account. Because, effectively, it was.

Scenario 2 — Planning a project kickoff (planning).

Help me plan the kickoff for our Q4 website redesign. Ask me clarifying
questions first — cover scope, people, risks, and timeline. Wait for answers.

The questions themselves are half the value. "Who has final sign-off on design?" is a question that, unasked, quietly destroys redesign projects around month three.

Scenario 3 — Research briefs (research). Before sending Claude off to analyze a market, a competitor, or a pile of survey feedback:

I want you to analyze this customer survey data for patterns. Before you start,
ask me clarifying questions about what decisions this analysis will feed into.

That last clause — what decisions this will feed into — is the difference between "here are some themes" and analysis you can act on.

Chain it into a repeatable workflow

This is my standard opener for any task over ten minutes. The chain looks like:

1. State the task in one messy sentence. Don't polish it — that's the point.
2. Add: Ask me clarifying questions before answering. Wait for my answers.
3. Answer the questions in lazy shorthand. Fragments are fine. The model doesn't care about your grammar.
4. Add: Summarize your understanding of the task in 3 bullets before you begin. — this catches the last remaining misunderstandings.
5. Let it run.

Steps 2 and 4 add about ninety seconds. They routinely save me from fifteen minutes of "no, that's not what I meant" revision loops. I've packaged an aggressive interview-style version of this as my Grill Me skill — it cross-examines you about a plan until there's nothing vague left.

Variations

• Ask me exactly 5 questions, most important first. — when you're in a hurry and want the model to prioritize.
• Ask me one question at a time. — better for thinking through genuinely fuzzy problems; the conversation becomes an interview.
• List the assumptions you'd otherwise make, and let me correct them. — same effect, faster, good for small tasks.
• Interview me like a consultant scoping a project, then write the brief. — my favorite for kicking off anything big.

Prompt 3: "Show me 3 examples first"

Before writing the full version, show me 3 different examples/approaches
in rough form. I'll pick a direction, then you'll develop it.

Why it works

When you ask for one draft, you get the statistically safest draft — the average of everything the model has seen. Averages are, by definition, unremarkable. Worse, once a full polished draft exists, you anchor on it. You start editing that draft instead of asking whether it was the right draft at all.

Asking for three rough options first does what good creative directors do: it separates choosing a direction from executing a direction. Humans are mediocre at generating options but excellent at picking between them. This prompt puts you in the judging seat, where you're strong, and keeps the model in the generating seat, where it's strong. It's the cheapest possible way to inject your taste into AI output — and taste is exactly what generic AI writing is missing.

Three workplace scenarios

Scenario 1 — Subject lines and openings (email).

I'm announcing a price increase to existing customers. Show me 3 different
opening paragraphs: one direct, one value-first, one personal. Don't write
the full email yet.

Before: one full email, opening with "We're writing to inform you of an upcoming change to your subscription." Delete. After: three openings side by side. The value-first one is clearly right for your audience. "Develop option 2, keep it under 150 words." Two minutes, and the hardest 20% of the email — the part people actually read — got real creative consideration.

Scenario 2 — Meeting agendas (meetings).

I'm running a 60-minute retro on a project that went badly. Show me 3 different
agenda structures — one blameless-postmortem style, one focused on the top 3
failures, one forward-looking. One paragraph each.

You'll immediately feel which one fits your team's mood this week. That feeling is the value — you can't get it from a single take-it-or-leave-it agenda.

Scenario 3 — Naming and framing (planning). Naming a project, an internal initiative, a newsletter section, a job title:

We're launching an internal program pairing senior staff with new hires.
Give me 3 naming directions with 3 names each: one warm/human, one
professional/corporate, one short/punchy. One-line rationale per direction.

Nine names, three philosophies, ninety seconds. Compare that to a meeting where six people stare at a whiteboard.

Chain it into a repeatable workflow

I call this the fan-out / pick / develop loop, and it's my default for anything where quality is subjective:

1. Show me 3 rough approaches. Label them A, B, C. Keep each under 100 words.
2. Pick one — or combine: Take B's structure with A's tone.
3. Develop it fully.
4. Now show me 2 variations of just the weakest section.
5. Repeat step 4 on any part that isn't landing.

Notice step 4: the fan-out works at every zoom level. Three whole drafts, then three versions of one paragraph, then three versions of one sentence. You're never stuck editing; you're always choosing. This pairs beautifully with the clarifying-questions prompt — questions first, then options. That two-step opening is baked into my Claude workflow template if you want the full system.

Variations

Variation	When to use it
"3 options: safe, bold, and weird"	You suspect the safe answer is wrong
"3 options from 3 perspectives: customer, CFO, skeptic"	Persuasive writing
"5 headlines, tell me which YOU think is strongest and why"	You want the model's judgment as a data point, not a decision
"2 options max, one sentence each"	Quick decisions, no ceremony

One warning: don't ask for ten options. Past four or five, quality degrades into filler and your choosing brain overloads. Three is the sweet spot.

Prompt 4: "Remember this — save it as a reusable skill"

Save this as a reusable instruction: [the thing you keep repeating].
Apply it in all our future conversations.

In Claude this lives in Settings as personal preferences, in Projects as project instructions, and in Claude Code as CLAUDE.md files and skills. In ChatGPT it's Custom Instructions and Memory ("Remember that I…"). In Gemini it's Saved Info. Same idea everywhere: stop re-typing your context every session.

Why it works

Here's the dirty secret of AI chat tools: by default, every new conversation starts from zero. The model has no idea who you are, what you do, or that you told it yesterday to stop writing "I hope this email finds you well." Most people re-explain their context dozens of times a week and don't even notice the tax they're paying.

Memory features flip a one-time effort into a permanent upgrade. Every instruction you save is a correction you never make again. Power users aren't better at writing prompts in the moment — they've accumulated a private layer of saved instructions that makes every prompt work better by default. That layer compounds. Six months in, their "vanilla" results are better than your carefully-prompted ones, and it looks like magic. It's not magic; it's savings.

Three workplace scenarios

Scenario 1 — Your writing voice (email). The third time you find yourself telling the AI to be less formal, save it permanently:

Save this as a standing instruction: When I ask for emails, write at an
8th-grade reading level, no corporate filler phrases ("I hope this finds
you well", "per my last email", "circling back"), short paragraphs,
one clear ask per email, sign off with just "Alex".

Before: every email draft starts 40% too formal and you edit it down, every time. After: drafts arrive pre-fitted to your voice. You've deleted a recurring 5-minute task from your life.

Scenario 2 — Your meeting-notes format (meetings). If you process meeting transcripts weekly, define the output format once:

Save this as my "meeting notes" skill: whenever I paste a transcript, return
(1) a 3-sentence summary, (2) decisions made, (3) action items as a table
with owner and deadline, (4) open questions. Flag any action item without
a clear owner in bold.

Now "meeting notes" is a two-word command instead of a paragraph of instructions. This exact pattern is why I built my meeting-notes-to-actions skill — it's the packaged version of that instruction.

Scenario 3 — Your standing context (planning/research). Save the facts you repeat constantly:

Remember: I'm a marketing manager at a 40-person B2B software company.
Our customers are mid-size logistics firms. My audience for most documents
is non-technical executives. When I ask for plans, always include a
"risks" section and assume a small budget.

Every future plan, analysis, and draft now starts from your reality instead of a generic one.

Chain it into a repeatable workflow

The habit that makes this compound is what I call the Rule of Three: the third time you type the same correction or context, stop and save it. Then, once a week (mine is Friday, 10 minutes):

• Skim the week's AI conversations for corrections you made more than once
• Turn each repeated correction into a saved instruction or memory
• Re-read existing saved instructions; delete any that no longer fit
• Promote your best repeated workflow (not just correction) into a named skill or project

That last step is the big one. A saved instruction fixes tone; a saved skill packages a whole multi-step process — my chains from prompts 1–3 all live as skills now. If you're curious how far this goes, my Claude Code explained for normal people post shows how the same idea becomes full automation, and the context engineering skill covers how to decide what belongs in permanent memory versus a single chat. Bonus: good saved context also means shorter prompts, which is one of the tricks in how I cut my Claude token usage in half.

Variations

• What do you currently remember about me and my preferences? — audit your memory layer; you'll be surprised what's in there.
• Forget the instruction about X. — pruning matters as much as saving.
• Based on our conversations, what standing instructions would improve my results? — let the model propose its own memory. Genuinely effective.
• Turn the process we just used into a reusable step-by-step skill I can invoke by name. — the workflow-capture move.

Prompt 5: "What are my blind spots here?"

Here's my plan/draft/decision. What are my blind spots? What am I not
seeing, what would a skeptic attack, and what's most likely to fail?

Why it works

By default, AI assistants are agreeable. Show them your plan and they'll tell you it's solid, add two mild suggestions, and wish you luck. This isn't lying, exactly — it's that "be supportive of the user's framing" is the path of least resistance, and your framing already contains your blind spots. If you don't explicitly ask for opposition, you get an echo.

The blind-spot prompt gives the model permission to disagree with you — and it turns out models are remarkably good critics when invited. They've absorbed every post-mortem, failure story, and "lessons learned" document ever written. They know how plans like yours have died before. All that knowledge sits unused until you ask the question that unlocks it.

There's also a psychological trick embedded here: asking "what are my blind spots?" is easier on the ego than hearing criticism from a colleague. It's a rehearsal space. You get to meet the strongest objections in private, before your boss or your client raises them in public.

Three workplace scenarios

Scenario 1 — Before you hit send (email). Any email where the stakes are real — a negotiation, a pushback, delivering bad news:

Here's an email I'm about to send declining a client's request for a discount.
What are my blind spots? How could this be read in a worse way than I intend?

Before: you send it; the client reads "we value your business, but..." as dismissive; two days of cleanup. After: Claude points out that your second paragraph implies their business isn't important enough for special treatment — which you didn't mean but absolutely wrote. One sentence changed. Crisis never happens.

Scenario 2 — The plan review (planning).

Here's our Q4 launch plan. Play three critics: a skeptical CFO, an
overloaded engineer on the team, and our biggest competitor. What does
each one see that I don't?

The CFO persona questions your revenue assumption. The engineer persona notices you scheduled the launch the same week as the platform migration. The competitor persona tells you exactly how they'd counter-program your announcement. Three meetings' worth of pushback, absorbed in five minutes, while the plan is still cheap to change.

Scenario 3 — Pressure-testing research (research/meetings). Before presenting findings or a recommendation:

I'm recommending we switch vendors based on this analysis. What are my
blind spots? What data would change my conclusion? What will the smartest
person in the room ask that I can't currently answer?

That last question is a superpower for meeting prep. Walking in with answers to the three hardest questions is what "being the most prepared person in the room" actually looks like.

Chain it into a repeatable workflow

I run a red-team pass on anything important before it leaves my hands. The chain:

1. What are my blind spots here?
2. Rank those risks by (a) likelihood and (b) damage. Which single one should I fix first?
3. Steelman the strongest objection — argue it as well as a smart opponent would.
4. Now help me fix the top 2 issues. Keep everything else unchanged.
5. Given the fixes, what's your remaining honest concern?

Note step 4's ending — "keep everything else unchanged" — otherwise the model will helpfully rewrite things that were fine. Make this chain a calendar habit: every deliverable gets a red-team pass the day before the deadline, not the hour before. This mindset comes straight from my Karpathy guidelines skill — treat the model as a brilliant but overeager collaborator whose confidence needs adversarial checking, including checking your own.

Variations

Variation	When to use it
"Argue against this as strongly as you can"	You're too in love with the idea
"What would make this fail in 6 months, quietly?"	Slow-failure risks nobody flags in reviews
"You are a skeptical [CFO/lawyer/customer]. React to this."	Rehearsing a specific audience
"Grade this A–F on logic, evidence, and clarity. Justify the grade."	Forcing specificity instead of vague praise
"What question am I not asking?"	Early-stage thinking, before there's even a plan

Prompt 6: "Don't stop until [explicit definition of done]"

Don't stop until [X]. Specifically, "done" means: [criterion 1],
[criterion 2], [criterion 3]. Before you finish, check your work
against each criterion and fix anything that fails.

Filled in, it looks like this:

Rewrite this report summary. Don't stop until it's under 200 words, every
claim has a number attached, there is zero jargon a new hire wouldn't know,
and it ends with one clear recommendation. Check each criterion before
you finish and fix any failures.

Why it works

AI models are sprinters with a bias toward plausible-looking completion. Give them a big task and they'll produce something that has the shape of an answer — the right sections, reasonable length — and stop, whether or not the job is actually finished. Ask for 20 ideas, get 12. Ask for a full review, get the first half reviewed carefully and the rest skimmed. The model isn't lazy; it just has no idea where your finish line is, so it draws one wherever the output starts looking complete.

A definition of done replaces the model's imaginary finish line with your real one. Even better, the "check your work against each criterion" clause makes the model audit itself before responding — and self-checking catches a shocking number of misses. This is the same principle that makes checklists work for surgeons and pilots: explicit criteria beat gut-feel completeness every time.

This prompt is also the foundation of every agent workflow. The reason tools like Claude Code can work for twenty minutes unattended is that they run against explicit completion criteria — tests pass, build succeeds. You're borrowing that discipline for everyday chat.

Three workplace scenarios

Scenario 1 — The inbox purge (email).

Here are 9 emails I need to answer. Don't stop until every single one has
a complete draft reply. "Done" means: 9 drafts, each under 120 words, each
with a clear next step, none starting with "Thanks for reaching out."
Number them 1–9 so I can check.

Before: you ask for help with your inbox, get beautiful drafts for the first four emails and a "...and similar approaches would work for the rest!" for the other five. After: nine numbered drafts. The numbering matters — it makes completion verifiable at a glance, for both of you.

Scenario 2 — Meeting follow-through (meetings).

Here's the transcript of today's planning meeting. Don't stop until every
decision and every action item is captured. "Done" means: each action has
an owner and a date, anything ambiguous is listed under "needs clarification"
rather than silently dropped, and you've done a second pass over the
transcript to catch items you missed the first time.

That "second pass" instruction alone typically surfaces two or three action items the first pass missed. Dropped action items are how meetings fail; this closes the leak.

Scenario 3 — Research that's actually thorough (research).

Compare these 4 project management tools for a 15-person team. Don't stop
until: all 4 are scored on the same 6 criteria in one table, pricing is
calculated for exactly 15 seats, each tool has a "worst thing about it"
listed, and you end with one recommendation and the strongest argument
against that recommendation.

Without the definition of done, you get a fluffy overview. With it, you get a decision document. Same model, same effort from you — completely different artifact. (For big research jobs, I've formalized this into my deep research brief skill, which is essentially a fill-in-the-blanks definition of done for research.)

Chain it into a repeatable workflow

Definition of done is the closer — it's the last thing I add after the other prompts have done their jobs. Which brings us to the full stack.

The complete workflow, chaining all six prompts:

1. Scope — state the task messily, then: Ask me clarifying questions first. (Prompt 2)
2. Direction — Show me 3 rough approaches before committing. Pick one. (Prompt 3)
3. Execute — Develop it. Don't stop until [definition of done]. (Prompt 6)
4. Pressure-test — What are my blind spots in this? Fix the top issues. (Prompt 5)
5. Translate — Explain the final result like I'm 10 — if you can't follow the simple version, something's wrong with the complicated one. (Prompt 1)
6. Capture — Save this whole process as a reusable skill named [X]. (Prompt 4)

Run that loop on one real task this week — a proposal, a difficult email, a plan. It takes maybe twenty minutes longer than the lazy way. The output difference is not subtle. And because step 6 saves the whole thing, the second run costs you almost nothing.

Variations

• Ask for 20, deliver 20. Count them before you respond. — the counting instruction fixes most "asked for N, got fewer" problems.
• If you can't meet a criterion, say so explicitly instead of lowering the bar. — prevents silent quality downgrades.
• Rate your output against each criterion 1–10 and improve anything below 8. — a built-in revision loop in one sentence.
• Continue from where you stopped. The definition of done has not been met. — the follow-up when a long task cuts off early.

Common mistakes (and how to fix them)

I've watched a lot of smart people use these prompts and quietly sabotage them. The failure modes are consistent:

1. Using the prompts once and declaring victory. The prompts aren't spells; they're habits. "Ask me clarifying questions" used once is a neat demo. Used on every non-trivial task for a month, it rewires how you delegate — to AI and, you'll notice, to people. Fix: pick two prompts, use them daily for two weeks, then add the rest.

2. Answering clarifying questions with a novel. The model asks five questions; people write five paragraphs. Shorthand is fine: "1: budget ~$5k. 2: audience is execs. 3: Friday. 4: formal-ish. 5: no." Speed keeps the habit alive. Fix: answer like you're texting.

3. Asking for options, then not choosing. Some people respond to three options with "can you combine all three and also add two more?" — which is just refusing to make a decision. The value of the fan-out is your judgment. Fix: force yourself to say "Option B" within thirty seconds. You can always change your mind later.

4. Hoarding memory until it rots. Saved instructions from your old job, your old role, your abandoned side project — stale memory actively degrades results, because the model keeps honoring rules that no longer apply. Fix: the Friday 10-minute review from prompt 4. Pruning is part of the practice.

5. Softening the blind-spot prompt. "Any small suggestions?" gets you small suggestions. If you ask gently, the model criticizes gently, and gentle criticism misses the fatal flaw. Fix: use strong framing — skeptic, attack, fail, steelman — and then (this is the hard part) actually change something based on the answer. A critique you read and ignore is theater.

6. Vague definitions of done. "Make it good and complete" is not a definition of done — it's the model's default target restated. If a colleague couldn't check your criteria with a yes/no for each, neither can the model. Fix: numbers, lists, and checkable conditions. "Under 200 words" beats "concise." "All 9 emails" beats "all of them."

7. Stacking all six prompts onto a 30-second task. Asking clarifying questions before "make this sentence shorter" is ceremony, not rigor. Fix: match the tool to the stakes. Quick tasks get quick prompts. The full six-step chain is for work with your name on it.

Your one-week challenge

Here's the checklist I'd hand you if we were sitting together:

• Monday: Add "Ask me clarifying questions before answering" to your first real task of the day. Answer in shorthand.
• Tuesday: Take something confusing from your inbox and run "Explain it like I'm 10" on it.
• Wednesday: Use "Show me 3 examples first" on one piece of writing. Choose within 30 seconds.
• Thursday: Run "What are my blind spots here?" on something before you send it. Change at least one thing.
• Friday: Save your three most-repeated instructions to memory. Delete any stale ones.
• Next big task: Write an explicit definition of done before you prompt. Watch what happens.
• End of week: Chain all six on one real deliverable using the workflow from prompt 6.

That's the whole system. No API keys, no paid courses, no plugins. Six sentences, used deliberately, on a schedule.

The people who seem "scarily good at AI" aren't running secret models. They've just stopped treating the text box like a search bar and started treating it like a capable colleague — one who needs a briefing, offers options, accepts criticism, remembers preferences, and works to an explicit standard. Those five behaviors are these six prompts. Now they're yours.

Resources & links

Keep going with these guides and free skills from the site:

• The Complete Prompt Engineering Guide (2026) — the full method behind everything in this post, from basic prompts to multi-step systems.
• My Claude Workflow Template — the exact reusable template I use to turn any recurring task into a Claude workflow.
• Claude Code Explained for Normal People — what Claude Code is and why non-developers should care, in plain English.
• How I Cut My Claude Token Usage in Half — practical tricks for getting more out of your usage limits.
• Secret Claude Code Commands — the lesser-known commands that unlock Claude's power-user features.
• Context Engineering skill — decide what belongs in permanent memory versus a single chat, and structure context so the model actually uses it.
• Grill Me skill — an interview-style version of the clarifying-questions prompt that cross-examines your plans until nothing vague remains.
• Karpathy Guidelines skill — working principles for treating AI as a brilliant but overeager collaborator that needs verification.
• Deep Research Brief skill — a fill-in-the-blanks definition of done for research tasks, so you get decision documents instead of fluffy overviews.
• Meeting Notes to Actions skill — paste a transcript, get decisions, owners, deadlines, and open questions in a consistent format.
• Superpowers skill — a bundle of my most-used reusable workflows, ready to install and invoke by name.

If this was useful, I share one practical AI workflow like this every week in my newsletter. Join here: https://alexandrrich.substack.com.