Lesson 5 of 32 · Part One, the principles
The Secret of Silent Suggestion
Few-Shot Prompting
Instead of describing the output you want, show two or three finished examples and let the model work out the pattern. A worked example carries decisions that adjectives cannot - how long an answer should run, where the line falls between two categories, which words you would never use. It is the difference between asking for a short caption and handing over three captions you consider short.
How to use The Secret of Silent Suggestion in ChatGPT
Lay the examples out as a plainly repeated pattern - the same labels in the same order every time - and leave the final label empty so the only job left is to complete it. Add one line saying what the examples have in common, because the pattern you care about (two lines, no hashtags, the material named) is not always the one it would infer on its own. For a sorting or extraction job you will do again tomorrow, the examples belong inside a Custom GPT rather than in your clipboard.
- Keep the label order identical in every example.
- State in one line what the examples have in common.
- Include the edge case you expect it to get wrong.
From GenAI for Everyone: Prompt Engineering Handbook, First Edition 2025, by Raja Rajeshwari Seshadri and Shreeram Iyer.
Lay the examples out as a plainly repeated pattern - the same labels in the same order every time - and leave the final label empty so the only job left is to complete it. Add one line saying what the examples have in common, because the pattern you care about (two lines, no hashtags, the material named) is not always the one it would infer on its own. For a sorting or extraction job you will do again tomorrow, the examples belong inside a Custom GPT rather than in your clipboard.
From the book Save your examples inside a GPT so it always follows your pattern. Upload a file of examples or past captions instead of pasting them each time, and use Regenerate to see another batch in the same pattern.
- Keep the label order identical in every example.
- State in one line what the examples have in common.
- Include the edge case you expect it to get wrong.
- Say which categories or words are not allowed.
- Move a recurring pattern into a Custom GPT.
Instead of
Give me science questions for practice.
ChatGPT version
I sit the Class 10 board paper in Science shortly. The two filled blocks below are my pattern. Study them, then complete the empty block the same way - same labels, same sequence, same tightness. --- Q: Why is a copper vessel coated with tin? (2 marks) A: Tin prevents copper reacting with acidic food, which could form harmful compounds. --- Q: Why do we see a rainbow after rain? (2 marks) A: Water droplets split sunlight into its colours by dispersion. --- Q: A: --- Produce five such blocks, drawn from chemical reactions and from light, split roughly evenly between the two. Every A runs to two lines - measure yours against mine before you hand them over, and shorten anything that has grown. Make the fourth block deliberately harder than either of mine, so I discover where my limit sits. Nothing may appear above, below or between the blocks: no title, no numbering, no closing remark. Count them, then reply.
Then loop it
- Good - the pattern held. Now three further blocks in the same shape but worth 3 marks, where each A runs to three lines rather than two. Keep the Q: and A: labels untouched.
- Now withhold the answers. Put the eight questions to me one at a time, wait for my reply, mark it out of the marks shown, and tell me what the model answer carried that mine did not. Afterwards, Regenerate if you want to show me a second set in the same pattern.
Why it worksTwo sample Q-and-A pairs teach the exact format, length and marks pattern better than any description.
Instead of
Write captions for my tailoring posts.
ChatGPT version
Learn my voice from these three captions of mine that did well. Do not describe the pattern back to me - just continue it. 1. "Warm wool, soft embroidery - a pheran made for Srinagar winters. Orders open till 10 Dec." 2. "Mother's old saree, new life as a kurta set. Bring yours, we'll remake it." 3. "Bridal suit, 40 hours of handwork. Every stitch, a blessing." What these have in common, and you must match all of it: two short parts, split by a dash or a full stop; the material or the labour named concretely; one line of warmth and one practical line; under 20 words; no hashtags; no emojis; no "DM us for prices". Now write captions for these three posts: - a kids' pheran - a lehenga alteration - a festive dupatta Give me two options for each, labelled A and B, where A sits closer to caption 1's style and B closer to caption 2's. Nothing else on the page - no explanation.
Then loop it
- Caption two of the alteration set is perfect. Keep that voice exactly. Write four more like it, all for alterations - an old saree, an old suit, a mother's lehenga, a loose kurta.
- Now add two Kashmiri or Urdu words to each caption, gently - the words my customers use themselves, not decoration. Still no hashtags. Then save all my examples inside a Custom GPT called "My captions" so I never paste them again.
Why it worksHer own best captions silently set the rhythm, length and emotion, so new ones sound like her.
Instead of
Sort this feedback from the relief camp.
ChatGPT version
A two-pass job. Pass one is the pattern; pass two is my data. **Pass one - the pattern.** These three are already labelled correctly: "No drinking water since yesterday" -> Water | High "Thank you for the blankets" -> Praise | Low "Children need notebooks to restart school" -> Education | Medium Read off them: an unmet need right now is High; a genuine need that survives a few days is Medium; gratitude, and everything that is not a request, is Low. **The closed lists.** Category takes one of Water, Food, Health, Shelter, Education, Praise, Other. Urgency takes one of High, Medium, Low. Invent neither. **Pass two - my data.** For every line I paste, emit one row of a markdown table whose columns are Comment | Category | Urgency | Notes. Preserve my ordering. Where a line carries a second need, that second one goes in Notes. Where nothing on the closed list fits, use Other and say in Notes why - never bend a category to avoid Other. Emit the table alone: no preamble, no totals, no advice. The names were stripped before I pasted. [Paste comments here]
Then loop it
- Now count how many fall in each category and list the High-urgency rows first, keeping the same four columns. Do not re-label anything you have already sorted - just reorder and count.
- Now a five-line summary for our coordinator to read out in the team meeting: what the numbers say, the most urgent thing, and what we should do first. Then save the categories and the three examples in a Custom GPT called "Camp feedback sorter" so tomorrow's batch takes one paste.
Why it worksThree labelled examples show the exact categories and format, so hundreds of comments get sorted consistently.
Questions
What is The Secret of Silent Suggestion in prompt engineering?
Instead of describing the output you want, show two or three finished examples and let the model work out the pattern. A worked example carries decisions that adjectives cannot - how long an answer should run, where the line falls between two categories, which words you would never use. It is the difference between asking for a short caption and handing over three captions you consider short.
When should I use this technique?
When you already have something you did well - old exam answers, your three best captions, a few sorted feedback forms - and you want the next fifty to come out the same way.
How do I apply it in ChatGPT?
Lay the examples out as a plainly repeated pattern - the same labels in the same order every time - and leave the final label empty so the only job left is to complete it. Add one line saying what the examples have in common, because the pattern you care about (two lines, no hashtags, the material named) is not always the one it would infer on its own. For a sorting or extraction job you will do again tomorrow, the examples belong inside a Custom GPT rather than in your clipboard.
What are the limits of this technique in ChatGPT?
Examples that all lean the same way tilt everything that follows - three warm, sentimental captions and every new one comes back sentimental. Put the awkward case in deliberately: the comment that fits no category, the alteration nobody feels anything about.
How do I improve the answer after the first try?
After the first answer, check whether the AI truly copied your pattern - length, tone, format, labels - or drifted. If it drifted, point to the gap: "Your answers are longer than my examples." If your examples are too alike, the output will be too, so add a different example and re-run. When the results are right, save your examples as a reusable template; next time, paste them and only change the new items.
Loop it
After the first answer, check whether the AI truly copied your pattern - length, tone, format, labels - or drifted. If it drifted, point to the gap: "Your answers are longer than my examples." If your examples are too alike, the output will be too, so add a different example and re-run. When the results are right, save your examples as a reusable template; next time, paste them and only change the new items.
This lesson on the other tools
The technique is the same; the way you write it is not. Claude, Gemini.Perplexity.