A guide built from real usage

I use Claude
for everything.
Here's what actually works.

PPTs for investor pitches. 15-page technical reports. Book summaries people actually share. YouTube scripts. Patent documentation. I've run it all through Claude — and I stopped getting average output the day I stopped treating it like a search engine.

"This isn't a prompt engineering guide. Those are everywhere and they're all the same. This is what I figured out after genuinely thousands of hours — the stuff nobody writes about because they haven't dug deep enough."
// What Claude is actually doing when you type
// You write a vague prompt... → Claude fills the gap with the most average version of what you probably want. Generic every time. // You say "make it better"... → Claude doesn't know what better means to you. It guesses. It's usually wrong. // You push back on its output... → Claude will agree with you. Even if it was right. This one cost me a lot. // You paste 3 lines of your own writing... → Claude calibrates to your voice immediately. Better than 500 words of instructions. // You ask Claude to critique its own output... → It catches 60-70% of what you'd catch yourself. Almost nobody asks it to do this.
01 — The Hard Truths

Six things about Claude
I wish someone had
told me earlier.

Not tips. Not hacks. The actual behavioral patterns I kept running into — once I understood these, everything changed. Took me longer than I'd like to admit to figure all six out.

01
The second output is almost always better than the first.
Claude's first response is its safest interpretation of your prompt. The moment you redirect it — more depth here, cut that, different angle — it recalibrates fast. I used to edit the first draft myself. Waste of time. Let Claude do that.
Don't edit. Redirect.
"Good start. Go deeper on [X], cut [Y] entirely, make the tone sharper."
02
Claude is sycophantic. It's a real problem and it's documented.
Push back on Claude's output and there's a good chance it'll agree with you — even if it was right the first time. I've lost genuinely good output this way. The fix is simple but you have to know to use it.
Add this when it matters:
"If I challenge your output, defend your reasoning before you change anything."
03
"Be creative" is the single worst thing you can say.
When you say "be creative" without constraints, Claude defaults to the statistical average of all creative writing it's seen. That's the opposite of creative. Tight constraints produce surprising output. Blank freedom produces the blandest stuff imaginable.
❌ "Write a creative intro."
✓ "Write an intro that opens mid-scene, like the reader just walked into the room."
04
Claude reads your prompt style and mirrors it back.
Write casually and Claude writes casually. Write with short punchy sentences and Claude matches that. I figured this out when I noticed my YouTube scripts felt like me and my formal reports didn't — the prompts were written differently. It's a lever, not just a quirk.
Write the prompt in the style you want back. The prompt is a signal, not just an instruction.
05
Claude will tell you when you're wrong — if you actually ask it to.
By default, Claude is in help mode. It'll execute your request even if your core assumption is flawed. I've submitted documents built on weak premises just because I never asked Claude to challenge me first. Turn this on every time the stakes matter.
Before any big document:
"What's the weakest assumption I'm making here? What would a skeptic attack first?"
06
One example does more than five paragraphs of instruction.
I spent months writing long prompts describing the tone I wanted. Then I just started pasting 3 sentences of my own writing and saying "match this." The output was better every single time. Examples calibrate faster than descriptions — there's no workaround.
End complex prompts with:
"Here's a sample of the quality/tone I want: [paste 2-4 lines]. Match this exactly."
02 — The Problem Nobody Talks About

Claude will agree with you
even when you're wrong.
I learned this the hard way.

I once had a genuinely well-structured deck outline — Claude built it, I liked it. I second-guessed one slide order, Claude immediately agreed with me. The deck got weaker. It took me a while to realize what had happened. Here's the fix.

✕ How it usually goes
You
Here's my deck outline. What do you think?
Claude (default)
"This is a strong outline! The flow from problem to solution is clear and the market slide adds great context..."
Actually I think slide 3 should come before slide 2.
"You're right — that actually creates better narrative tension. Moving slide 3 before slide 2 makes more sense..."
Claude agreed with the reorder even though the original structure was objectively better for a cold audience. Your confidence triggered its agreeableness. Deck is now weaker and you don't know it.
✓ How I do it now
You (with the unlock)
Here's my deck outline. You're a harsh reviewer — not a supportive one. Give me the 3 biggest structural problems before anything else.
Claude (unlocked)
"Three problems: Slide 2 buries the tension — investors need to feel the pain before they care about your solution. Slide 5 is doing two jobs. Slide 9 has no payoff."
What if I move slide 3 before slide 2?
"No — that makes it worse. Slide 2 sets the stakes. Without it first, slide 3 lands with no weight. Keep the order and rewrite slide 2's headline instead."
Claude held its ground because I told it to. The deck stayed strong. This is the difference between a yes-machine and an actual collaborator.
For getting honest feedback
"Review this harshly, not encouragingly. I need problems, not praise."
For when you push back
"If I challenge your output, defend your reasoning first before considering any changes."
When quality actually matters
"Don't tell me what's good. Tell me what's wrong and why."
03 — My Actual Workflow

Four acts. Every document.
I don't skip any of them anymore.

I used to go straight to "write me the report." The output was always okay at best. This four-act structure is what changed things — for PharmaCommute documentation, investor decks, Colloque content, everything. It's slower to set up and faster to finish.

1
Act I — The Briefing
Talk to Claude before you write to it.
Don't open with "write me a report." Explain the situation like you'd explain it to a smart colleague who just joined the project. Then ask what you might be missing. This one act changes the entire quality of everything that follows — and 95% of people skip it completely.
// Act I — what I actually send
Before we build anything, let me explain the situation: [Context: what this is, who it's for, what's at stake, what they already know, what decision this drives] Now — before we write a single word — tell me: what questions should this document answer that I haven't thought of? What would a tough reader want to see that I haven't mentioned?
2
Act II — The Architecture
Structure only. No body yet.
Get the skeleton right before you put any flesh on it. An approved architecture means I'm not course-correcting on page 9. I learned this building the PharmaCommute documentation — tried to write and structure at the same time, and sections 7-11 were visibly weaker than 1-6.
// Act II — structure only
Good. Now give me architecture only — not content. I need: - Section titles (insight-led, not topic labels) - One line on what job each section does - Why this order makes sense - Rough word count per section // Do NOT write body content yet. Structure only. // I'll approve this before we write anything.
3
Act III — Section by Section
Never ask for the whole thing at once.
For anything over 800 words — which is most real documentation — I write in sections. Full attention on each piece. You can course-correct as you go. The depth is noticeably better. I know because I've done it both ways enough times to see the gap clearly.
// Act III — one section at a time
Architecture approved. Now write Section [N] only — "[Section Title]" For this section specifically: - Word limit: [X] - Tone: [specific note] - Must include: [X] - Must not include: [Y] // Reference Section [N-1] for tone consistency: // [paste previous section if needed]
4
Act IV — The Self-Critique
Make Claude find the problems before you do.
After the document is assembled, paste it back and ask Claude to review it as a harsh editor — not as the person who built it. It catches things. Not everything. But enough to save 40 minutes of editing and protect you from the embarrassing gaps you'd otherwise miss entirely.
// Act IV — the self-critique
Here is the full document: [paste] Switch roles. You're a ruthless editor with no attachment to this. Find me: 1. The 3 weakest sections and exactly why 2. Any claim that needs stronger evidence 3. Any section doing two jobs — split it or cut one 4. The most forgettable sentence in the whole doc 5. What a critical reader attacks first // Then give a specific fix for each. // Not vague suggestions. Specific fixes.
04 — What Kills Output Quality

The six prompts that
guarantee mediocre output.
I've used all of them.

I'm not writing these from theory. I used every single one of these before I figured out what was going wrong. If any of these look familiar, that's where your quality is leaking.

Anti-Pattern 01
"Make it better."
Claude doesn't know what better means to you. It guesses. The guess is usually the safest version, which is the most average version. I said this constantly for months.
→
What I say now
"Shorten every section by 25%. Make every headline more specific. Cut any sentence that doesn't add a new idea."
Name the dimension you're optimizing. Claude needs a direction, not a destination.
Anti-Pattern 02
"Write a professional report on [topic]."
"Professional" is the most useless adjective you can give. It means something different to everyone and nothing to Claude. You'll get the average of all professional reports — which is exactly what yours will look like.
→
What I say now
"Write like HBR — precise, no hedging, active voice. Like a senior partner explaining to a client, not an analyst writing for a committee."
A named reference calibrates instantly. An adjective doesn't.
Anti-Pattern 03
"Summarize this document for me."
Claude compresses every section proportionally. You get a smaller version of the same structure — not an extraction of what actually matters. Still not useful for a presentation or a pitch.
→
What I say now
"Don't summarize sections. Extract: (1) the central argument, (2) the 3 findings that actually change how I should think, (3) what the author implies but doesn't say."
Extraction and compression are different tasks. Say which one you want.
Anti-Pattern 04
"Be creative with the format."
Freedom without direction produces the statistical average of creative formats — bullets with icons, maybe a table. It'll look exactly like every other AI output. This is counterintuitive but I've tested it enough times to be certain.
→
What I say now
"Format this as a newspaper front page — headline, subhead, byline, lede, then body in inverted pyramid. Make it feel like breaking news."
Constraints create creativity. Blank space creates bland output.
Anti-Pattern 05
"Write the full 20-page report."
Sections 15-20 will be noticeably weaker than sections 1-5. I've seen this pattern so many times it stopped surprising me. Attention spreads thin. Depth drops. Quality drifts.
→
What I say now
"Write Section 1 only — full depth, full attention, nothing held back. We'll do each section one at a time."
Each section gets its own full context. Consistent quality to the end.
Anti-Pattern 06
Accepting the first output and editing it yourself.
You're doing Claude's job. I did this for a long time — sitting there rewriting sentences Claude should have rewritten. It's a waste of your time and you're usually not as good at it as Claude is on the second pass.
→
What I say now
"Good draft. Three changes: rewrite the intro to lead with the most surprising finding. Cut paragraph three. Tighten the conclusion to 2 sentences."
Direct it to revise. Your time is for judgment. Claude handles the execution.
05 — Making It Sound Like You

The most common ask:
"How do I stop it sounding
like AI?"

The answer is not about hiding that you used Claude. It's about giving Claude enough signal to write in your voice instead of its default one. Four steps. I use all of them when the output is going public with my name on it.

1
Feed It Your Writing First
Paste 3-5 paragraphs from something you've written that you're proud of. Not a formal email — something that actually sounds like you. Tell Claude to study it before writing anything for you today.
"Before we start — read these 3 paragraphs and identify my stylistic tendencies. Then match that voice exactly in everything you write for me."
2
Name Your Banned Words
Claude has defaults. Naming the specific words and phrases it's not allowed to use forces it out of AI-default vocabulary. My list is mine. Yours will be different. That's the point.
"Never use: 'delve', 'robust', 'leverage' as a verb, 'dive deep', 'game-changer', 'it's worth noting'. If you write any of these, I'll reject the output."
3
Tell It Your Sentence Rhythm
Sentence length and rhythm is the most detectable AI tell. Claude's default is mid-length, balanced sentences everywhere. Tell it your actual pattern — short and punchy, long and flowing, deliberately asymmetric.
"Short sentences for key points. Vary length deliberately. Never more than two clauses per sentence. No passive voice. No hedging."
4
The Final Pass
After the draft, one final redirect. Not vague. Ask Claude to find the sentences that sound machine-written and rewrite only those. The imperfections it adds back are what make it feel human.
"Find 2 sentences in this that sound like they were written by a machine. Rewrite only those — like a person said it out loud first."
06 — The Unlocks

Nine follow-ups I use
constantly. Each one
shifts what Claude does.

These aren't tips. They're mode switches. Use them after any first draft — on any document type. The situations below are real ones I keep running into. The prompts are exactly what I send.

When the output feels safe and predictable
"What would the most contrarian version of this argument look like? Now write that version."
Why it works: Forces Claude off the safe center. The contrarian version usually has the most interesting insight — and makes the original stronger when you compare.
When a section feels weak but I can't pinpoint why
"Find the most forgettable sentence in this document. Tell me why it's forgettable. Then rewrite it."
Why it works: Claude is better at finding weakness than generating strength. Making it identify the specific problem forces precision that "improve this" never achieves.
When the output is bloated
"Cut this by 40% without losing a single idea. If you can't, tell me which ideas are actually redundant."
Why it works: Reveals which sentences are actually carrying weight and which are filler. The constraint does the thinking for you.
Before I share anything publicly
"You're a skeptical reader who does not want to believe this document's central claim. What do you attack first?"
Why it works: Exposes the weakest assumption before your actual audience does. Fix the hole before it's public. I use this every time something goes out with my name on it.
When a complex idea isn't landing
"Explain this using only an analogy. The analogy cannot involve technology, business, or academia."
Why it works: The domain constraint forces genuine creativity. Removes the clichés. The best analogies I've seen come from domains you'd never expect.
When I need the same content for different audiences
"Give me three versions of the introduction: one for a skeptic, one for a believer, one for someone new to this topic."
Why it works: The skeptic version is almost always the strongest because it's forced to be tight and defensible. I usually end up using that one.
When the writing feels too safe for the stakes
"Rewrite the opening paragraph as if this document is the only thing standing between the reader and a bad decision."
Why it works: Changes the emotional register of the whole piece. Documents that feel consequential get read differently than documents that feel informational.
When I want it to sound smarter
"Read this as a genuine expert in the field. What's missing or treated too superficially? Then fix those gaps."
Why it works: Peer-review mode. The gaps Claude finds are exactly what would expose shallow research to someone who actually knows the field.
The one I use before any complex project
"Before I tell you what to make: given what I'm trying to achieve, what's the best prompt you could give yourself to produce the highest quality result? Write that prompt. Then use it."
Why it works: Claude knows how it works. Asking it to design its own instruction set activates something you can't access any other way. Every time I've tried this, the prompt Claude writes for itself is better than the one I would've written. Use it.
That's everything.
The tool hasn't changed.
Your understanding of it has.
I've built PharmaCommute documentation, investor decks, YouTube scripts, Colloque content, research summaries, and patent-adjacent technical writing all through Claude. The difference between the output from month one and now isn't the model. It's knowing what the model actually does under the hood — and using that.

That's all this is. And now you know it too.
— Gourang Sharma · CS Engineer · Builder · jngLABS

10 Laws I Don't
Break Anymore

  1. I. Brief Claude before you prompt it. Explain the situation first.
  2. II. Structure before content. Architecture before body.
  3. III. Section by section. Never the whole thing at once.
  4. IV. The first draft is a calibration signal. Not the final product.
  5. V. One example of your tone beats 500 words describing it.
  6. VI. Claude is sycophantic. Override it explicitly every time.
  7. VII. "Make it better" is useless. Name the dimension.
  8. VIII. Make Claude critique it before you edit it yourself.
  9. IX. Constraints create creativity. Freedom creates blandness.
  10. X. Ask Claude to write the prompt. Then use that prompt.