12 Prompt Shortcuts

What this is for

You already write decent prompts. The problem is what comes back: fake-cheerful copy, shallow answers, or the AI agreeing with a bad idea because you proposed it. These are modifiers, not full prompts: one line each, attached to whatever you're already asking. Works in ChatGPT or Claude, either one.

The 12 shortcuts

1. Kill the AI voice: strips "delve," "unlock," and upbeat-consultant tone.

Rewrite this in a plain, direct human voice. Cut throat-clearing openers and adjective stacking. Short sentences. No exclamation points.

Before: "In today's dynamic landscape, it's crucial to unlock your team's potential!" → After: "Your team is underused. Here's where."

2. Explain it simply: plain-English pass, checks itself for jargon.

Explain this like you're talking to a smart colleague outside this field. Define any technical term in the same sentence. Reread and cut anything too dense.

Before: "Leverage the API's idempotency guarantees." → After: "This makes sure a duplicate request only counts once."

3. Think before answering: forces reasoning before the answer, not pattern-matched speed.

Before answering, list the three most likely ways this could go wrong or be misread. Then give your answer, adjusted to avoid those three problems.

Before: a generic five-step plan → After: the same plan minus two steps that assumed a budget you don't have.

4. Stop agreeing with me: turns off sycophancy, forces the counter-case.

Don't tell me I'm right. Assume my plan has a flaw and find it. Argue the strongest case against what I just proposed.

Before: "Great idea, that should work well!" → After: "Your pricing punishes your best customers for buying more. Here's why."

5. Give me options: forces real range, not three versions of one idea.

Give three genuinely different versions of this. Label the approach each one takes in five words or less.

Before: one generic email → After: a direct ask, a story-led ask, a data-led ask, clearly labeled.

6. Ask first: stops the model guessing at missing context.

Before doing anything, ask me the 3 questions you'd most need answered to do this well. Wait for my answers.

Before: a proposal built on wrong assumptions → After: three sharp questions on budget, audience, deadline, then a proposal that fits.

7. Show sources or say so: forces an admission instead of invention.

Only state facts you're confident are accurate. Say "I'm not certain" instead of guessing. Name a source if you'd normally cite one; if you can't, say so.

Before: a specific stat stated as fact → After: "I don't have a reliable figure. Treat this as a rough estimate."

8. Sound like me: matches your actual writing sample.

Here is a writing sample of mine: [PASTE 200-300 WORDS OF YOUR OWN WRITING]. Match this voice (sentence length, tone, formality) for everything you write me from here on.

Before: generic polished tone → After: your actual cadence, including your habit of opening with the conclusion.

9. Cut it in half: real compression, not filler-trimming.

Cut this to half the word count without losing substance. Cut whole sentences, not words within them. Cut the weakest point, not the shortest sentence.

Before: a 400-word update with three examples → After: a 200-word update with the one example that matters.

10. Name your assumptions: surfaces the guesses baked into the answer.

List every assumption you made to produce this answer, in plain language, before showing the answer itself.

Before: a budget plan with no explanation → After: "I assumed a $50K budget and a team of 4," then the plan.

11. Grade yourself: forces a self-critique pass before shipping.

Rate this answer 1-10 against the goal I gave you. Say specifically what's weak. Rewrite it to fix those weaknesses.

Before: a flat first draft → After: "6/10, too generic in the middle," followed by a sharper draft.

12. Put it in a table: forces scannable structure for comparisons.

Format your answer as a table. Decide which columns matter most for comparing these, then build it.

Before: three paragraphs comparing vendors → After: a table with rows per vendor, columns for price, setup time, lock-in risk.

Where this goes wrong

• People paste all 12 at once. Stacking them muddies the instruction. Use one or two per prompt, matched to the actual problem.
• People expect the sycophancy-killer (#4) to stay off after one use. It doesn't. The model drifts back to agreeable by your next message; reapply when you need pushback.
• People treat "cite or say you don't know" (#7) as a fact-check. It isn't. The model still can't verify anything live; it's just less likely to bluff. Verify what matters yourself.
• Some shortcuts pull against each other. #4 (argue with me) and #8 (sound like me) make an odd pair if your own writing voice is agreeable, and #9 (cut it in half) will happily cut the assumptions #10 just surfaced. When two are in play, tell the model which one wins.

Save your top four so you stop retyping them

The four most people keep: voice match (#8), sycophancy-killer (#4), self-critique (#11), and table format (#12). Here they are assembled as one block, ready to paste into a custom-instructions field:

Standing instructions for how you respond to me:
1. Match my voice: [PASTE 200-300 WORDS OF YOUR OWN WRITING]. Match
   this sentence length, tone, and formality in everything you write.
2. Don't tell me I'm right by default. When I propose a plan, find
   its strongest flaw before agreeing.
3. Before finishing any substantial answer, rate it 1-10 against
   what I asked for, name the weakest part, and fix it.
4. When comparing three or more things, use a table.

Where to save it, ChatGPT: Settings, then Personalization, then Custom Instructions; paste the block under "Anything else ChatGPT should know."

Where to save it, Claude: for always-on behavior, put it in your account's personal preferences (in Settings, under your profile). A Project's instructions only apply inside that project, which is the right scope for work-specific shortcuts but not for "everywhere" ones. For the voice-matching half specifically, Claude's Styles feature (in the style picker in the chat input) is purpose-built: create a custom style from your writing samples once and select it whenever you want your voice.

The 2-minute version

Pick the shortcut matching today's problem, paste it above or below your real prompt, send. Then save the assembled block above into custom instructions once, and numbers 4, 8, 11, and 12 become your defaults. These modifiers sharpen every tool in The AI Stack; the two are companions.

One honest note: there's no hidden switch or secret API these "codes" unlock. They're plain-English instructions the model reads like any other text. They work because most people never specify what they actually want, not because of any trick.

---

Want to know where you actually stand with AI? The free AI Readiness Assessment takes 10 minutes and gives you a personalized roadmap.

Weekly, on Tuesdays. The Gen Xcelerator: practical AI intelligence for people who already know how to do the job. Subscribe free.

Experience Is the API. GenXcelerate