Prompting / Prompt Champs

20 Prompt Tricks That Put You Ahead of 99%

07:57 runtimePublished June 25, 202620 key ideas

Wrap pasted text in delimiters like triple quotes or XML tags, so the model never mistakes the source material for new instructions. This blocks prompt injection from inside your data. Try: "Summarize the text in…

Watch on YouTube

The complete breakdown

Use this page with the video

Watch the episode above for the visual explanation, then use the notes below to revisit each idea, example, and practical move.

01

Fence off your content

Wrap pasted text in delimiters like triple quotes or XML tags, so the model never mistakes the source material for new instructions. This blocks prompt injection from inside your data. Try: "Summarize the text in <doc>...</doc>."

Put pasted text in <doc>...</doc>

02

Make it echo the task

Before it answers, ask it to restate your request in its own words. This surfaces misreadings while they're cheap to fix instead of after a wrong wall of text. Add: "First repeat back what I'm asking, then wait for my OK."

"Restate the task, then wait"

03

Name the exact shape

Replace vague words like "short" with hard counts: "exactly 5 bullets, max 8 words each." Models hit precise targets far better than fuzzy ones because there's nothing to guess. Specify rows, sentences, or word caps explicitly.

"Exactly 5 bullets, 8 words max"

04

Show what NOT to do

Don't just describe the ideal output — paste a bad example and label why it fails. Negative examples fence off failure modes that positive ones can't reach. Try: "Avoid openings like this: 'In today's fast-paced world...'"

Paste a 'bad example' to ban it

05

Set the audience

State exactly who's reading: "explain to a CFO" lands very differently than "explain to a 10-year-old." The model tunes vocabulary, depth, and what it assumes you know. Naming reading level beats just asking for "simple."

"Explain this for a CFO"

06

Answer from this only

Paste your reference text and add: "Answer only using the text above; if it's not there, say so." This grounds replies in your material instead of the model's memory, and forces an honest "not found" over a guess.

"Only use the text above"

07

Demand 'according to'

Tell it to phrase claims as "according to the document..." This anchors every assertion to a source and makes unsupported statements feel out of place to the model, so it invents less. Each line now points back to real text.

"According to the source..."

08

Ask it to flag doubt

Add: "Mark anything you're unsure about and rate your confidence." Models will happily state shaky claims as fact, so making uncertainty an explicit output turns a hidden risk into a visible flag you can verify.

"Mark what you're unsure of"

09

Give do's AND don'ts

List both what to include and what to avoid as explicit rules: "Use plain English. Don't use jargon, em dashes, or hype." Hard boundaries on both sides cut the drift you get when you only describe the good version.

Two lists: include / avoid

10

Ask what you missed

After the answer, follow with "What am I missing?" or "What would make this better?" A fresh pass with a critical lens catches gaps and angles the first attempt skipped — it's the cheapest quality upgrade you can run.

"What am I missing here?"

11

Three takes, then merge

Ask for three genuinely different approaches, then have it combine the strongest parts into one. Forcing variety beats settling for the first answer, and the synthesis step keeps the best of each. Try: "Give 3 options, then merge."

"3 distinct versions, then merge"

12

Outline before drafting

On big tasks, get a plan or outline first, approve it, then expand each section in follow-ups. Breaking work into steps keeps the model from losing the thread on long jobs and lets you steer before it commits to a draft.

Plan first, expand part by part

13

Make it grade itself

Have it define success criteria — a short rubric — then score its own answer against each one. Naming the bar before judging forces honest self-assessment and exposes where the draft falls short. Try: "Set 5 criteria, then grade 1-5."

"Build a rubric, then score it"

14

Step back first

Before the specific answer, ask for the underlying principle: "What general rule governs this?" Reasoning from the concept down leads to sounder answers than jumping straight to the detail. Then have it apply that rule to your case.

"What's the principle here?"

15

Stage an expert debate

Ask two experts who'd disagree to argue it out, then have the model weigh both and conclude. Surfacing the strongest case for each side exposes tradeoffs a single confident answer would paper over. Great for judgment calls.

Two experts debate, then conclude

16

Show it the hard case

When giving examples, include one tricky edge case, not just easy ones. Models pattern-match to your samples, so a clean-only set teaches it to fumble the messy inputs. One well-chosen exception sharpens the whole behavior.

Add one edge case to your examples

17

Set standing rules

Put recurring preferences in custom instructions or a system prompt — tone, format, what to skip. Then you stop retyping them every chat and every answer inherits them by default. Set it once: "Be concise. No preamble. Ask if unclear."

Custom instructions, set once

18

Critique then fix

Add: "List 3 weaknesses in your answer, then rewrite to fix them." Naming flaws before revising forces a real second pass instead of a cosmetic touch-up, and the model is often a sharp critic of its own work.

"3 weaknesses, then fix them"

19

Prefill the reply

Start its answer for it to lock the format: end your prompt with "Sure, here's the JSON:" or "{". The model continues from your opening, so it skips the chatty preamble and commits to the exact structure you began.

Begin its reply with "{"

20

Let it write the prompt

Before running a hard task, ask the AI to improve your prompt first: "What would make this prompt clearer? Rewrite it, then ask me anything missing." It knows what context it needs better than you do — meta-prompting beats guessing.

"Rewrite my prompt first"

More practical AI guides, without the hour-long lecture.

Subscribe on YouTube