Free to read · no account needed

9 min · prompt

Show 1–2 examples (few-shot) without confusing the model

Goal: Steady format with one or two short before/after examples—no long pastes, no secret facts taught to the model.

Best for: When the model “almost” matches the shape you want (tables, ticket replies, bullet style).

Few-shot means showing a small number of input→output examples. One or two excellent examples beat five messy ones. Label sections clearly so the model knows where the real task starts.

Few-shot means one to three short examples that show shape—not a novel of your past chats. Bad examples teach the model your bad habits.

Sanitize examples like you sanitize screenshots: no customer names, no keys, no internal URLs you would not post publicly.

Why beginners skip this

Pasting two random examples feels like few-shot—if they don’t match the goal, the model learns wrong. Pick same-class real examples first.

Step 1 — Rules for examples

  • Show format, not confidential content
  • Keep examples short
  • Label: EXAMPLE INPUT / EXAMPLE OUTPUT / NOW YOUR TURN
  • Prefer invented toy data over real customer PII

Step 2 — Copy pattern

Paste two-good/one-bad into the format prompt you already use—redact names and customer IDs first.

Follow this format exactly.

EXAMPLE INPUT:
[short]
EXAMPLE OUTPUT:
[short, perfect shape]

NOW YOUR TURN
INPUT:
[real content]
Boundaries: [no invented numbers]
Example

Filled prompt

I need subject lines for B2B docs emails. Provide 2 good examples and 1 bad example (too hype), then ask the model to write 5 new ones for: "Q3 usage report ready for finance." Rules: ≤6 words, no emoji, no “Use tiny before/after samples so format”

What good output looks like

  • Goal/Format/Boundaries all present for “Use tiny before/after samples so format sticks—wit”
  • One example beats stacked IMPORTANT words
  • Version/date kept when template changes
  • Pass/fail: zero clarifying turns on live task

Step 3 — When few-shot fails

If examples fight the boundaries: simplify, switch to a rubric + self-score, or open a new chat with one example only.

Example

Filled prompt

Task: rewrite support replies in our tone.
Provide 2 good examples + 1 bad (too apologetic + invents refund).
Then rewrite this ticket reply using the pattern. Ban new policies.

What good output looks like

  • One example beats stacked IMPORTANT words
  • Version/date kept when template changes
  • Pass/fail: zero clarifying turns on live task
  • Goal/Format/Boundaries all present for “Use tiny before/after samples so format sticks—wit”

Save this card

Card: One or two short examples steady the format—keep long text and secrets out of few-shots. Keep this beside the tool; read before you start.

Before you start

Keep the last failing output visible. You will repair the prompt for “Use tiny before/after samples so format” against that failure—not a fresh toy ask. Guide: prompt-few-shot-that-actually-helps. Keep the real job visible the whole time; a blank box means stop and fetch the artifact first.

How to use this week

This week for a real task pick 2 same-class good examples + 1 counterexample, mark what to learn, then run.

Two good, one bad

That shape teaches faster than ten long dumps. Sanitize every example.

Keep it real

Run this guide’s Practice now on work you will actually ship this week—not a fictional scenario. Dated notes beat perfect theory.

Ship the practice

Close the tab only after Practice now is done on a real artifact.

Do not bookmark-only

Finish Practice now, then tick this guide on your path. Titles unread are not skills.

Close the loop

Write one line: next time I paste examples with no learning marks, I refuse to run.

Example

Filled prompt

Job: turn raw bullets into Slack updates.
2 good exemplars + 1 bad (invents "on track 90%").
Then convert these bullets: [canary green | write-path lag at 18:00 | need freeze window].

What good output looks like

  • Version/date kept when template changes
  • Goal/Format/Boundaries all present for “Use tiny before/after samples so format sticks—wit”
  • Pass/fail: zero clarifying turns on live task
  • One example beats stacked IMPORTANT words

Red flag

Few-shots full of secrets or novels; five long examples—cut to one or two short format demos.

Pitfalls

PitfallFix
Five long examplesCap at 1–2
Examples full of PIIToy data
No “NOW YOUR TURN”Hard label
Teaching facts to memorizeTeach format only
Example

Filled prompt

Temptation: skip the habit for prompt-few-shot-that-actually-helps to save five minutes.
Job still: Use tiny before/after samples so format
Smallest correct pass: Done visible · checkable artifact · Unknowns [?].
Boundaries: draft only; no silent send; no invented metrics.
Return: artifact + refuse-line said aloud.

What good output looks like

  • Names the shortcut refused (specific to prompt-few-shot-that-actually-helps)
  • Still delivers checkable output for “Use tiny before/after samples so format”
  • Unknowns labeled; draft-only
  • Pass/fail: refuse-line dated (prompt-few-shot-that-actually-helps #4)

Practice now

Contrast pair: one good + one bad example only (Asia/Shanghai).

Few-shot drill: two tiny examples only; delete a third that adds noise (Asia/Shanghai).

Thin-spot drill: rewrite the Save card with real nouns from your work (repo, doc, inbox). Generic placeholders fail the drill.

Tonight for a real task pick 2 same-class good examples + 1 counterexample, mark what to learn, then run.

Today card (copy) — prompt-few-shot-that-actually-helps: 1. Artifact I will produce: ___ 2. Pass/fail check: ___ 3. Refuse-line if tempted to skip verify/send: ___ (date: Asia/Shanghai) 4. One never-do from this guide: ___

Extra pass: do the card on a real file today—reading alone does not count.

One sentence takeaway

2 good + 1 counterexample; mark what to learn, then run.

Next

GFB · Rubrics · Debug ignores

← Back to home · Learn · Tools · Compare · Start