6.18 · Providing examples — Few-shot prompting

Mở đầu: "Show, don't tell"

Ninh xương bò 2 giờ. Nêm gia vị vừa miệng. Rau thơm tươi. Phục vụ nóng.

Cách B — Show:

Ninh xương bò 2 giờ. Đây là ví dụ bát phở chuẩn: [Hình ảnh bát phở thật + ghi chú: nước dùng trong, hành lá thái mỏng, thịt bò tái chính giữa, chanh ớt bên cạnh] Chú ý: nước dùng phải trong (không vẩn đục), thịt thái mỏng 2mm, không dày.

Cách B cho bạn:

• Trực quan hơn
• Chi tiết ẩn (độ trong, độ mỏng) không tả được bằng lời
• Chuẩn cụ thể để so sánh kết quả của mình

Few-shot prompting = Cách B với Claude. Cho Claude ví dụ input/output → Claude học pattern.

Zero-shot vs Few-shot

Zero-shot

Chỉ instruction, không example:

Classify this tweet as positive or negative: "Yêu app này quá!"

Strength: Concise. Works cho task common. Weakness: Claude guess — có thể miss corner case.

One-shot

1 example:

Classify tweets as positive or negative.

Example:
Tweet: "Great game tonight!"
Label: Positive

Now classify:
Tweet: "Yêu app này quá!"

Strength: Claude biết format expected. Weakness: 1 example không cover nuance.

Few-shot (2-5 examples)

Nhiều example, đặc biệt corner case:

Classify tweets as positive or negative.

Examples:

Tweet: "Great game tonight!"
Label: Positive

Tweet: "Oh yeah, I really needed a flight delay tonight! Excellent!"
Label: Negative  (← sarcasm!)

Tweet: "The food was OK, nothing special."
Label: Negative  (← mild, not neutral)

Now classify:
Tweet: "Yêu app này quá!"

Strength

Cover corner cases (sarcasm, mild sentiment). Weakness: Tốn token, risk over-fit example.

Rule

2-5 examples thường đủ. > 5 chỉ cho task rất nuanced.

XML structure cho examples

Anthropic đặc biệt "quen" với XML tags. Structure examples clear:

prompt = """Classify tweet sentiment: Positive, Negative, or Neutral.

<examples>
<example>
<input>Great game tonight!</input>
<output>Positive</output>
</example>

<example>
<input>Oh yeah, I really needed a flight delay tonight! Excellent!</input>
<output>Negative</output>
<reasoning>Sarcasm. Words positive, but context shows frustration.</reasoning>
</example>

<example>
<input>The food was OK.</input>
<output>Neutral</output>
</example>
</examples>

Now classify:
<input>{tweet}</input>"""

Benefits:

• Claude dễ parse boundaries
• Có thể add <reasoning> để teach logic, không chỉ label
• Consistent structure → consistent output

Corner case examples

Ví dụ mạnh nhất là corner case — scenario Claude hay fail.

Sentiment analysis

Normal:

• "Love it!" → Positive
• "Terrible service" → Negative

Corner cases:

• "Not bad" → Positive (mild)
• "Could be worse" → Neutral/slightly Negative
• "It's fine, I guess" → Neutral
• "Best movie since Plan 9 from Outer Space" → Negative (sarcasm, Plan 9 is notorious bad movie)

Which to include? Corner cases. Skip obvious ones.

Product categorization

Normal:

• "iPhone 15" → Electronics
• "Running shoes" → Sports

Corner cases:

• "Fitness tracker" → Electronics OR Sports? → Pick 1, explain
• "Yoga mat" → Sports OR Home? → Pick 1
• "Water bottle" → Kitchen OR Sports? → Context matters

Case study: Meal planner với examples (v5)

Từ bài 6.17, score = 7.86. Add examples:

prompt_v5 = f"""Generate a one-day meal plan for an athlete.

<example>
<input>
Height: 180cm, Weight: 80kg, Goal: Build muscle, Restrictions: None
</input>
<output>
Total: 3200 calories | Protein 200g | Carbs 350g | Fat 80g

8:00 Breakfast (700 cal):
- 3 eggs scrambled (210g) — 210 cal, 18g protein
- Oatmeal with banana (80g dry oats, 1 banana) — 400 cal, 10g protein
- Greek yogurt (150g) — 90 cal, 15g protein

12:00 Lunch (900 cal):
- Grilled chicken breast (200g) — 330 cal, 60g protein
- Brown rice (80g dry) — 280 cal, 6g protein
- Broccoli + olive oil (150g + 10ml) — 80 + 90 cal

15:00 Snack (500 cal):
- Protein shake (30g whey + 250ml milk) — 250 cal, 30g protein
- Almonds (30g) — 180 cal, 6g protein

19:00 Dinner (1100 cal):
- Salmon (200g) — 400 cal, 40g protein
- Sweet potato (250g) — 215 cal, 4g protein
- Salad + avocado (200g + 100g) — 200 + 160 cal

Rationale: High protein (2.5g/kg body weight) for muscle building. 
Timing around workout (assumed 17:00-18:00).
</output>
</example>

Now generate for:
Height: {h}, Weight: {w}, Goal: {goal}, Restrictions: {restrictions}"""

Score v5: 8.9/10 (+13% vs v4)

Example cho:

• Exact format structure
• Level of detail (portions in g, timing, rationale)
• Tone (direct, numeric)

Sourcing examples từ evals

Best examples là từ previous evals — outputs scored highest:

Workflow

1. Chạy prompt v1 qua test dataset
2. Identify case scored 9-10/10
3. Extract input + output pairs
4. Use as examples in v2

best_cases = sorted(eval_results, key=lambda x: x.score, reverse=True)[:3]

examples_text = "\n\n".join(f"""
<example>
<input>{case.input}</input>
<output>{case.output}</output>
</example>""" for case in best_cases)

prompt_v2 = f"""{base_instruction}

<examples>
{examples_text}
</examples>

Now handle:
<input>{{input}}</input>"""

Insight: Prompt engineering cải thiện exponentially khi bạn có eval loop — vì bạn extract insight từ previous outputs.

Anti-patterns

Example quá generic

<example>
<input>Tweet about food</input>
<output>Positive</output>
</example>

Vấn đề

Input không cụ thể, Claude không học được gì.

Fix

Example input phải là thực tế cụ thể.

Example contradict instruction

Instruction: "Respond in 100 words."
<example>
<output>[3-word response]</output>
</example>

Vấn đề

Claude confused, follow example hay instruction?

Fix

Example must exemplify instruction, not contradict.

Biased examples

<example 1><input>good</input><output>Positive</output></example>
<example 2><input>great</input><output>Positive</output></example>
<example 3><input>amazing</input><output>Positive</output></example>

All positive → Claude có thể bias toward Positive for all future input.

Fix: Balance — include positive, negative, edge, neutral.

Over-specific examples

Examples quá match 1 scenario → Claude fail khi input khác scenario đó.

Fix: Diverse examples. 3 examples should cover 3 different scenarios.

Too many examples

10+ examples bloat prompt, tốn token, Claude overwhelm.

Fix: 2-5 examples. Pick diverse + corner case.

Áp dụng ngay

Bài tập 1: Upgrade prompt với 3 examples (20 phút)

Lấy prompt bạn đã có. Thêm 3 examples:

1. Normal case
2. Corner case
3. Edge case

Test: score tăng không? Nếu tăng → keep. Nếu giảm → examples có vấn đề (generic? contradict? biased?).

Bài tập 2: Example + reasoning (15 phút)

Mở rộng 1 example, thêm <reasoning>:

<example>
<input>...</input>
<output>...</output>
<reasoning>This is the right output because...</reasoning>
</example>

Reasoning giúp Claude học logic, không chỉ pattern.

Test: xem output của Claude có "smarter" không.

Tóm tắt bài học

Tài liệu tham khảo

• Anthropic: Use examples