Make it not look AI-made (set once)
Slides & TalksUsing AI
Set the fonts, layouts and phrases to avoid once, so slides and images stop looking AI-made.
Fill the blanks and it copies as you filled it
Whenever you make slides or images, always do this. - One font, Pretendard, for both headings and body - Vary how much text is on each slide. Some with two bullets, some with a single sentence - Mix at least three layouts. Don't repeat heading-plus-bullets on every slide - Mix the headings too — a question, a single word, a full sentence - One accent color, used decisively. No muddy grays or purple gradients - Don't use phrases like "in today's fast-changing world", "leveraging innovation", "unprecedented", "synergy", "paradigm" - Numbers should be concrete — "3,400 people", not "many people" - Two icons at most. Do the rest with whitespace and size differences - Either put the same footer on every slide or leave it off entirely
0 of 0 blanks filled · anything left blank copies with its brackets
Put this in your custom instructions or style settings once, not in every conversation.
What this prompt does
Compared with asking the same thing without the length and format lines.estimate
Estimated output tokens
700
Asked loosely
320
This prompt
Energy (Wh)
1.215
Asked loosely
1.215
This prompt
Token counts are estimated from character counts, not measured. Energy uses the per-band figures measured by Jegham et al. 2025 as published.
A shorter answer, but the same energy band. The measurements are per band, and this much does not cross one.
Reported backfire — Specifying output lengthSpecifying a length longer than needed backfires (a 34% increase in 11% of scenarios)
Reported backfire — Specifying output formatA 78% increase in one GPT-4o scenario. The paper classifies this one as apply-selectively
Where these figures come from
We read the prompt to see whether it carries that instruction. The percentages below are published figures for the rule itself, not a measurement of what this prompt saves.
Specifying output length
38.7%mean reduction · 95% CI 29.7–47.7%
Podder et al. 2026 BP#5 · 18 use cases / 57 prompts
by modelGPT-4o 35%Command R+ 49%Mistral-7B 32%
Reduced in every scenario. No backfire reported
Specifying output format
49%mean reduction · 95% CI 37–61%
Podder et al. 2026 BP#6 · 14 use cases / 45 prompts
by modelGPT-4o 40%Command R+ 50%Mistral-7B 57%