PromptSharp
Issue #001 · Wed, Jun 24, 2026
Sharpen how you instruct AI — from your first prompt to full agents. |
Few-shot calibration: teach the shape, not the answer
Two or three examples will out-perform two paragraphs of rules — if you pick the right examples. |
Few-shot prompting means showing the model a handful of input→output pairs before the real input. The mistake beginners make is using easy, obvious examples. The model already knows those. Your examples should cover the boundaries — the tricky case, the 'don't do this' case, the edge case — because that's where ambiguity lives. Also keep the format of every example byte-for-byte identical; the model copies structure as strongly as content. |
▸ Copy-paste prompt
Classify each support message as: BUG, BILLING, FEATURE_REQUEST, or OTHER.
Return only the label.
Examples:
Message: "The export button does nothing when I click it."
Label: BUG
Message: "I was charged twice this month."
Label: BILLING
Message: "Would love a dark mode someday."
Label: FEATURE_REQUEST
Message: "Your product is great, just saying hi!"
Label: OTHER
Message: "App is slow AND I think my invoice is wrong."
Label: BILLING
Now classify:
Message: "[paste message]"
Label: |
▸ Level it up
New to this: Give the AI 3 examples of what you want before your real ask. Match the format of each example exactly — same labels, same line breaks.
Vibe coder: Curate examples that include the confusing case (the last pair above is a deliberate two-topic tie-breaker). Edge-case examples teach 5x more than easy ones.
Advanced: Build a labeled gold set and rotate few-shot exemplars dynamically by retrieving the 3 nearest-neighbor cases to the live input (dynamic few-shot / k-NN selection). Track per-class accuracy to spot which exemplars actually move the needle. |
Today's takeaway: Few-shot teaches by demonstration — so demonstrate the hard cases, not the easy ones. |
|
|
|
Know someone who'd sharpen their prompts with this?
Forward it. Refer 3 friends and we'll send you our private prompt-pack.
Refer & get the prompt-pack → |
|
|
|
|