Skip to main content
$ai-systems-by-jon
Back to Guides
ClaudePromptingAI Tools

The 5 Words That Change How Claude Answers You

6 min read

Claude has a default voice and it is not the one you want

Ask it something real and it hedges. It softens. It opens with "great question" and closes with a summary nobody asked for. That is the default, and the default is tuned to be safe for everybody, which makes it sharp for nobody.

You change it with one word.

The honest part, first

These are prompt patterns, not official Anthropic commands. They work because of how Claude reads plain-language instructions, and that also means results shift a little between model versions. Most land exactly as described. A few need a nudge. Test the ones you plan to lean on before you build them into your daily work.

The five below, plus the format ones further down, hold up the most reliably across versions.

How they work

One word, dropped into your prompt. Claude reads it as an instruction about how to answer, not what to answer.

RAW: Explain how compound interest works.

Stack two with a plus sign:

RAW + PHD: Explain how compound interest works.

Some belong at the front of the prompt, some at the end. Each one below says which.

The five from the video

  1. RAW — front. Cuts every hedge, disclaimer, and softener. You get the answer with no cushion. Point it at something you are too close to, like whether your business idea is any good.
  2. BUILD — end. Skips the explaining and just builds the thing. Working code, working structure, working output. Put it after the request, not before.
  3. HUMAN — front or end. Rewrites the answer so it reads like a person typed it. Kills the polished AI cadence. Best on emails and captions.
  4. STEELMAN — front. Argues both sides at their strongest, then picks. This is the one for real decisions, like hiring someone full-time versus using a contractor.
  5. PHD — front or end. Answers like someone twenty years into the field instead of a summary generalist. More depth, more precision, fewer training wheels.

Making answers shorter or longer

CodeWhereWhat it does
TIGHTendThe absolute minimum without losing the point
TL;DRfrontTwo-sentence answer up top, detail underneath
BULLETSendBullet points instead of paragraphs
TABLEendReformats into a table
ONE-SENTENCEfrontHas to fit in one sentence. Kills waffling
NO-YAPfrontStrips filler, disclaimers, and closing wrap-ups
DEEPeitherThe opposite of TIGHT. Fully unpacks it
RECAPendBoils a long conversation down to one takeaway

Changing the tone

CodeWhat it does
CASUALConversational, like texting a friend
FORMALPolished and business-appropriate
WITTYDry humor without turning into a comedy bit
WARMSofter and more empathetic. Good for people you care about
DIRECTNo pleasantries. Instruction or answer only
ANTIFRAGILEStrips "might" and "could" and commits to a claim
NO-EMOJISBlocks emojis and decorative characters

Building and shipping

CodeWhereWhat it does
SHIPendThe finished product. Like BUILD but for anything, not just code
FIRST-DRAFTfrontRough pass. Go fast, do not polish
FINAL-DRAFTfrontPublish-ready, nothing left to clean up
PRODendProduction code with error handling and edge cases covered
MVPfrontThe smallest version that tests the idea
REMIXfrontFive different versions of what you already have

Thinking harder

CodeWhat it does
PROS-CONSStructured, with a weight on each side
RED-TEAMEvery way this fails or gets attacked
BLUE-TEAMThe best case for the idea
FIRST-PRINCIPLESReasons from the ground up, no inherited assumptions
SECOND-ORDERThe downstream effects, not the obvious one
INVERTSolves for the opposite. What would guarantee failure?
DECIDEForces a real recommendation. No "it depends"
WEIGHRanks the options by impact or effort

Checking the work

CodeWhat it does
FACT-CHECKVerifies every factual claim before writing it
CITEIncludes a source for each claim
VERIFYCritiques its own answer before showing you
STEEL-DOUBTArgues against what it just said and finds the weak link

Formatting for where it is going

CodeWhat it does
EMAILSubject line, greeting, body, sign-off
DMThree sentences, max
PITCHA sixty-second spoken pitch, first person
SLACKCasual, short paragraphs, minimal formatting
TWEETOne post under 280 characters
THREADA numbered thread of five to eight posts
HOOK-ONLYJust the opening line. Nothing else
CTAAdds a call to action to whatever you wrote
SCRIPTA spoken video script with the beats marked
CAPTIONA social caption with hook, body, and CTA

Dialing the depth

CodeWhat it does
ELI5Explains it like you are five. Kills jargon
ELI-EXPERTAssumes you know the basics and skips them

Stacking

Two codes is where this gets good. Combinations that earn their keep:

  • A real decision, fastSTEELMAN + DECIDE + TIGHT
  • Code you can actually shipPHD + PROD + NO-YAP
  • Copy that sounds like a personHUMAN + TIGHT + NO-EMOJIS
  • An email that gets a replyDIRECT + EMAIL + WARM
  • An opener for a videoHOOK-ONLY + CASUAL
  • Analysis with no hedgingPHD + RED-TEAM + ANTIFRAGILE
  • Learning something newELI5 + BULLETS + NO-YAP

Three things that trip people up

  • Putting the code in the wrong spotat the front, Claude treats it as a lens for the whole answer. At the end, it lands more like an edit pass. RAW, PHD, and STEELMAN want the front. BULLETS, TABLE, and CTA want the end.
  • Stacking too manythree or more and they start fighting each other for priority. Two clean codes beats four messy ones almost every time.
  • Assuming they are permanenteach model version reads them a little differently. If you switch models and your favorite stack goes flat, that is why. Re-test it.

One more place to put them

These work in system prompts too. If you are building an agent or a custom skill, drop them in as defaults instead of retyping them. NO-YAP and PROD as standing defaults give you a noticeably sharper coding assistant from the first message.

Frequently asked questions

  • Are these official Claude commands?No. They are prompt patterns that work because of how Claude reads plain-language instructions. Nothing here is documented by Anthropic.
  • Will they work on every model?Mostly. Behavior drifts a little between versions, so test any stack you depend on after switching models.
  • Where do I put the code?Front for anything that shapes the thinking, like RAW or STEELMAN. End for anything that shapes the format, like BULLETS or TABLE.
  • How many can I stack?Two is the sweet spot. Past that they compete and the answer gets muddy.
  • Do they work outside Claude?Many carry over to other assistants, since they are plain instructions rather than special syntax. They are written and tested for Claude.

Start here

Pick five and use them today. The thing that changes is not answer quality so much as how much time you stop spending cleaning up the output.