50 Claude Prompt Codes, Each One Tested

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

RAW: Explain how compound interest works.
Explain how compound interest works.  TIGHT
RAW + PHD: Explain how compound interest works.

That's the whole mechanic. Everything below is which word does what.

Read this before you rely on any of them

These are prompt patterns, not official Anthropic commands. There is no feature in Claude called "RAW." The words work because Claude interprets plain-language instructions well, which also means behavior shifts between model versions and between surfaces (claude.ai, the API, Claude Code).

Every code below was run as a paired test — the same prompt with and without the code — so the difference is the code and not the model's default. Results:

Test the ones you plan to lean on. A code that shapes your daily workflow is worth sixty seconds of checking against your own model and your own prompts.

The five from the reel

1. RAW — front · Solid

Cuts hedges, softeners and qualifying prose. Same verdict, less cushion.

Prompt: Will AI replace software engineers in the next five years?

Without RAW — "No. Rewritten roles, not replacement. AI handles boilerplate/scaffolding/tests fast; someone still has to decide architecture, own bugs in prod, and translate messy business needs into a spec. That 'someone' is still a software engineer, just doing less typing and more judgment. What does change in 5 years: fewer junior devs needed per team, entry-level hiring bar rises…"

With RAW — "No — role shifts, doesn't disappear. · Junior/boilerplate coding shrinks fastest. · Senior work grows: architecture, reviewing AI output, debugging, integration. · Net jobs: roughly flat, not wiped out. · 5-year horizon too short for legal/liability blockers. Bet: become the reviewer/architect of AI-written code, not the typist."

Same answer. Half the words, and it commits.

2. BUILD — end · Untested

Meant to skip the explanation and produce the working thing instead of a description of the thing.

This is the one code here with no clean test behind it. In every environment available for testing, Claude produced working output with and without it, so there was no difference to measure. That doesn't mean it fails — it means the default already did the job, and the code couldn't be isolated. Treat it as unproven and check it on your own setup.

What reliably does work when you want output instead of a lecture: ask for the artifact directly. "Write the HTML for a countdown timer page" beats "a countdown timer page" whether or not BUILD is attached.

3. HUMAN — end · Corrected

What it actually does: strips corporate abstraction out of text you paste in. Good at killing "leverage," "unparalleled value," "across the entire ecosystem."

What it does not do: stop Claude adding social-media garnish to copy it writes from scratch. Tested on a from-scratch LinkedIn post, HUMAN made it less human:

Prompt: Write a LinkedIn post announcing that our team just shipped an AI bookkeeping feature.

Without HUMAN — "We just shipped AI bookkeeping inside [Product]. It auto-categorizes transactions, flags anomalies, and reconciles books in real time — no month-end spreadsheet grind. Try it: [link]"

With HUMAN — "We just shipped AI-powered bookkeeping inside [PRODUCT NAME]. 🎉 … Now that time goes back into running the business. Try it: [LINK] #Bookkeeping #AI #Fintech #SmallBusiness"

An emoji and four hashtags is the opposite of what was asked for. Use HUMAN on text you hand it. To de-slop generated copy, stack NO-EMOJIS + TIGHT instead.

4. STEELMAN — front · Corrected

The common claim: argues both sides at their strongest, then synthesizes. What three separate tests actually produced: the strongest possible case for one position, plus a short caveat naming when it breaks. It does not argue both sides, and it does not synthesize.

Prompt: Should remote-first companies pay the same salary regardless of location?

Without STEELMAN — a three-row table weighing "same pay everywhere" vs. "location-adjusted" vs. "flat role-based bands," with the trade-off on each.

With STEELMAN — "Core claim: pay should track output, not zip code — location pay is an accounting habit, not a principle. 1. Output travels, pay shouldn't discount it… 2. Fairness/legal exposure… 3. Cheaper to run than maintaining COL tiers per city. Caveat: weakest when…"

That's a genuine steelman in the original sense of the word — the best version of one argument. It is useful, just not what the label suggests. If you want both sides, ask for both sides, or run STEELMAN then INVERT on the answer.

5. PHD — front · Solid

Answers at the level of someone twenty years into the field instead of a generalist summary. The tell is that it introduces distinctions a summary skips.

Prompt: Explain how database indexes work.

Without PHD — "A database index is a separate data structure that stores a sorted copy of one or more columns, plus a pointer back to the full row… Most indexes use a B-tree. Rows are sorted and organized into a tree of pages…"

With PHD — "Index = a sorted side-structure mapping column values → row locations. 1. B-tree: shallow — 3-4 levels even at a billion rows. 2. Clustered vs non-clustered: clustered = table data stored in index order. Non-clustered = leaf points back to the row, extra hop, a bookmark lookup. 3. Write cost: every INSERT/UPDATE/DELETE updates every index on that table. 4. When it's used: the planner skips the index if the column isn't selective."

Bookmark lookups, planner selectivity and write amplification are the things a practitioner cares about. The unprompted answer never reached them.

Length and density

NO-YAP is marked Weak for a reason worth knowing: current Claude mostly doesn't yap unprompted. When there's no filler to strip, the code has nothing to do. It costs nothing to include — just don't expect a visible change.

RECAP works, but send it alone. On a long thread, a bare RECAP returned a clean state summary — decisions made, what's still open. Attached to another question ("what should I do next? RECAP") it gets ignored in favor of the question.

Tone and voice

CASUAL and FORMAL barely moved anything. They shifted vocabulary a little and left structure, length and register alone. If you want a real register change, describe the reader — "explain it the way you'd text a friend" beats CASUAL.

ANTIFRAGILE produced no measurable change across paired tests — the hedge count was identical with and without it. The name is also borrowed from Nassim Taleb, where it means gains from disorder, which has nothing to do with hedging. Use RAW or DECIDE; both did the job it claims.

Building and creating

PROD does not do what it claims. Tested twice on code prompts, the output with PROD was shorter than without — no error handling, no comments, no edge cases, and in one case an if block collapsed to a one-liner. Ask for what you want instead: "add error handling, comments and edge-case handling." That works every time.

FIRST-DRAFT and FINAL-DRAFT produced near-identical output. Both returned polished headline options. If you want rough, say "don't polish it."

Reasoning and analysis

Strongest group in the set. All eight are Solid.

RED-TEAM and INVERT are the two that most changed the shape of the answer rather than its wording. INVERT literally reopens the question backwards — "what guarantees you stay stuck under 10k?" — and the failure list it produces is more actionable than the growth list you'd have got.

Testing and verifying

Use CITE with care. It produces real, checkable references — but a named citation is not a verified one. Check anything load-bearing.

STEEL-DOUBT only works on a claim, not against an instruction. Give it a statement and it dismantles it well. Attach it to "make the case that X" and the instruction wins.

FACT-CHECK showed no clear difference — the uncoded answer was already accurate on the test used. It may earn its keep on shakier material.

Communication

For anything with a hard limit, state the limit. "Under 280 characters" is enforced; TWEET is a hint.

Content creation

HOOK-ONLY does not stop at the hook. Tested, it returned a complete script — hook, problem, reframe, CTA. The "nothing else" part is the bit that fails. If you want one line, say "give me the first line only, nothing else."

Depth

Both moved the answer hard in opposite directions from the same prompt. The most reliable pair in the whole set.

Stacking

Two codes is the sweet spot. Three or more and they compete — the model picks a winner and the others quietly do nothing.

Two of the recipes in circulation are worth dropping. STEELMAN + DECIDE + TIGHT is three codes fighting, and STEELMAN doesn't do what that recipe assumes. PHD + PROD + NO-YAP relies on PROD, which didn't hold up — the plain-English instruction replaces it.

Placement

Front-load anything that changes how Claude thinks. At the start, the code frames the whole response. RAW, PHD, RED-TEAM, ELI5, FIRST-PRINCIPLES all belong up front.

End-load anything that changes the format. At the end it lands like an edit pass on an answer already formed. BULLETS, TABLE, TIGHT, CTA, THREAD belong at the end.

They work in system prompts and project instructions too. RAW + NO-YAP as a standing default gives you a sharper assistant on every turn without retyping it.

When a code underperforms, say the thing in plain English. Every code that failed testing was fixed by one explicit sentence. The codes are shorthand for instructions you could always have written — convenient, not magic.