Bài tập tổng hợp — Viết prompt production-grade

Đề bài

Build 1 prompt production cho scenario bạn chọn từ 3 option dưới. Iterate v1 → v4, đo improvement.

Option A: Resume screener

Input: CV text + job description.

Output JSON:

{
  "match_score": 85,
  "strengths": ["...", "...", "..."],
  "gaps": ["...", "..."],
  "recommendation": "Interview | Pass | Junior pool",
  "questions_to_ask": ["...", "..."]
}

Option B: Customer email reply generator

Input: Incoming customer email + context (product, previous interactions, team policy).

Output: Draft reply email (plain text).

Option C: Meeting notes → action items

Input: Meeting transcript (50-500 words).

Output JSON:

{
  "decisions": ["...", "..."],
  "action_items": [
    {"task": "...", "owner": "...", "deadline": "..."}
  ],
  "risks_raised": ["..."],
  "followup_needed": true
}

Quy trình

Bước 1: Set goal (5 phút)

Trong notebook 04_final_prompt.ipynb, write:

Bước 2: v1 — Baseline (10 phút)

Viết prompt đơn giản nhất. Chạy qua 5 test case. Chấm điểm bằng tay (1-10):

## Goal

App: [Option A/B/C]

Prompt mục tiêu: [1 câu tóm tắt]

Success criteria (checklist khi eval):
- [ ] [criteria 1]
- [ ] [criteria 2]
- [ ] [criteria 3]
- [ ] [criteria 4]
- [ ] [criteria 5]

Test dataset: [bao nhiêu cases, source, cover edge không?]

Bước 2: v1 — Baseline (10 phút)

Bước 3: v2 — Apply Clear & Direct (10 phút)

Rewrite first line thành action command. Add basic structure.

## v1 — Baseline

Prompt:
[paste]

Test results:
- Case 1: score 3/10. Issue: ...
- Case 2: score 2/10. Issue: ...
- ...

Average: 2.6/10

Bước 3: v2 — Apply Clear & Direct (10 phút)

Bước 4: v3 — Add Guidelines (10 phút)

Add 5-6 quality guidelines. Specify format, length, must-include.

## v2 — Clear & Direct

Prompt:
[paste]

Changes from v1:
- Line 1: changed from "X" to "Y"

Test results:
- Case 1: score 5/10. Better structure.
- ...

Average: 4.5/10 (+73%)

Bước 4: v3 — Add Guidelines (10 phút)

Bước 5: v4 — Add Examples + XML (15 phút)

Wrap context trong XML tags. Add 2 examples (1 normal, 1 edge case).

## v3 — + Guidelines

Prompt:
[paste]

Changes:
- Added 6 guidelines

Test results:
- Case 1: score 7/10. All required fields present.
- ...

Average: 7.2/10

Bước 5: v4 — Add Examples + XML (15 phút)

Bước 6: Verify & Document (10 phút)

Check:

• [ ] Run 5 test case lần 2 với v4 — stable score?
• [ ] Có regression (case trước đây pass giờ fail) không?
• [ ] Document changelog trong file prompts_changelog.md

## v4 — + Examples + XML

Prompt:
[paste]

Changes:
- Wrapped input in XML tags
- Added 2 examples

Test results:
- Case 1: score 9/10. Edge case handled.
- ...

Average: 8.5/10

Deliverable

Submit 3 thứ:

1. File my_prompt.py

2. File prompts_changelog.md

# my_prompt.py

PROMPT_V4 = """[final prompt]"""


def run_prompt(inputs: dict) -> str:
    # ... code gọi Claude với prompt
    pass

2. File prompts_changelog.md

3. Notebook 04_final_prompt.ipynb

Chạy v4 qua 5-10 test case, print output. Ready to demo.

# Prompt Evolution

## v4 (Current, 2026-04-20)
Score: 8.5/10
Changes: Added examples + XML structure
Author: [your name]

## v3 (2026-04-19)
Score: 7.2/10
Changes: Added 6 quality guidelines

## v2 (2026-04-19)
Score: 4.5/10
Changes: Clear & direct rewrite

## v1 (2026-04-18) — Baseline
Score: 2.6/10

Reference: Ví dụ hoàn chỉnh (Option A — Resume Screener)

Version này score 8.5/10 trên test 10 case. Ship production.

PROMPT_V4 = """Evaluate candidate resume against job description. Return structured assessment.

<job_description>
{job_description}
</job_description>

<candidate_resume>
{resume}
</candidate_resume>

<evaluation_framework>
1. Technical skills match (weight: 40%)
2. Experience relevance (weight: 30%)
3. Cultural fit signals (weight: 15%)
4. Growth potential (weight: 15%)
</evaluation_framework>

<guidelines>
- match_score: integer 0-100 (weighted avg of above)
- strengths: exactly 3 items, specific quotes from resume
- gaps: 1-3 items, concrete missing skills/experience
- recommendation: "Interview" if score > 70, "Pass" if < 50, "Junior pool" if 50-70
- questions_to_ask: exactly 3 items, target gaps or ambiguity
</guidelines>

<examples>
<example>
<input_job>Senior Python Engineer - 5+ years - AWS + Kubernetes</input_job>
<input_resume>John - 6 years Python, Django, some AWS basics, no K8s</input_resume>
<output>
{{
  "match_score": 72,
  "strengths": [
    "6 years Python exceeds 5-year requirement",
    "Django framework implies strong backend fundamentals",
    "Some AWS exposure"
  ],
  "gaps": [
    "No Kubernetes experience mentioned",
    "Unclear depth of AWS (only 'basics')"
  ],
  "recommendation": "Interview",
  "questions_to_ask": [
    "Can you walk through an AWS project in detail?",
    "Any exposure to container orchestration (Docker Compose, ECS, K8s)?",
    "Most complex distributed system you've designed?"
  ]
}}
</output>
</example>

<example>
<input_job>Senior Data Scientist - ML + deep learning</input_job>
<input_resume>Mary - Business analyst 3 years, Excel, SQL, some Python</input_resume>
<output>
{{
  "match_score": 25,
  "strengths": [
    "SQL skills transferable to data engineering",
    "Some Python exposure shows willingness to learn"
  ],
  "gaps": [
    "No ML or deep learning experience",
    "No statistical modeling background",
    "Role is too senior for current skill set"
  ],
  "recommendation": "Pass",
  "questions_to_ask": [
    "Interest in ML training path?",
    "Would Junior Data Analyst role interest you?",
    "Any ML projects outside of work?"
  ]
}}
</output>
</example>
</examples>

Return JSON only, no markdown wrapper."""

Tips & Tricks

Tip 1: Debug output quality by adding reasoning field

Tạm thời thêm <reasoning> vào output để thấy Claude "nghĩ" gì:

Sau khi hiểu logic, remove _reasoning cho production.

Tip 2: Test với extreme cases

Prompt production phải handle gracefully.

Tip 3: Version control prompt

Commit prompt vào git cùng code. PR review cho prompt như PR review code.

• Input rỗng
• Input quá dài (10K words)
• Input ngôn ngữ khác (Vietnamese, Chinese)
• Input ambiguous

{
  "match_score": 72,
  "_reasoning": "...",  // ← debugging
  "strengths": [...]
}

Tip 3: Version control prompt

git add my_prompt.py prompts_changelog.md
git commit -m "feat: prompt v4 — add examples for edge cases"

Self-review checklist

Content quality

Code quality

Documentation

• [ ] v4 cover tất cả success criteria?
• [ ] Test dataset đa dạng (normal + edge)?
• [ ] 3-5 round iteration?
• [ ] Score tăng consistently?
• [ ] my_prompt.py có docstring?
• [ ] run_prompt() có type hints?
• [ ] Error handling (API fail, JSON parse)?
• [ ] prompts_changelog.md có tất cả version?
• [ ] Notebook demo chạy được fresh?
• [ ] Test case cả positive và negative?

Tóm tắt

🎯 4 techniques áp dụng theo thứ tự: Clear & Direct → Specific Guidelines → Examples → XML.

🎯 Mỗi iteration đo score, không phải "feeling".

🎯 Changelog + version control — prompt như code.

🎯 Test dataset đa dạng — 20% edge case.

🎯 Ship khi score stable 8+/10 qua 3 lần run.