Overview

You want performance reviews to sound a certain way. Specific accomplishments, balanced (good and growth areas), direct, honest. Not corporate fluff. Not harsh. Not generic.

You could describe your style to AI: "Be specific, balanced, honest, direct. Include accomplishments and growth areas. Avoid corporate language."

That helps a little. You get something closer to what you want.

Or you could paste three performance reviews you've written that you're proud of. You say: "These are examples of our company's reviews. Draft a similar review for [new person]."

The second approach works infinitely better. AI learns your exact style, length, structure, tone, and voice from examples faster than from descriptions. Much faster. It's the difference between "I described my preferences" and "you saw my actual work."

This is few-shot prompting: teaching AI by example instead of description.

This lesson teaches you how to use few-shot prompting for every HR task that benefits from style consistency. You'll learn what makes a good example. You'll learn how to annotate examples so AI understands why they work. You'll learn to build example libraries that become your style standards.

Why This Matters for HR Professionals

Few-shot prompting saves you rewriting time. Instead of telling AI to "be less corporate," you show it corporate-free reviews. It gets the pattern. The AI has learned your voice from seeing real examples.

It also ensures consistency. If all your reviews are built from the same 3-4 examples, they'll have similar voice, length, structure, and balance. Your reviews look like they come from the same company, with the same values. Consistency matters.

It also makes your communication feel like you. Your recruiting emails have your tone. Your policy clarifications have your clarity. Your difficult conversations have your balance of honesty and care.

Few-shot prompting is the difference between generic AI output and company-specific, values-aligned output. It's the difference between "this could have come from anywhere" and "this is clearly from our company."

What Makes a Good Example

For few-shot prompting to work, your examples need to be:

1. Genuinely Good
The examples should be work you're proud of. Work that exemplifies what you want more of. If you're not proud of an example, don't use it as a template for more output. Your examples are your standard. They need to be good.

2. Representative
Show different scenarios. If you're doing performance reviews, show: a high performer, a middle performer, someone with growth areas. If you're doing recruiting emails, show: reaching out to a passive candidate, sourcing from a referral, recruiting from a different industry.

Variety teaches AI the pattern. Sameness bores it and limits what it learns.

3. Annotated
Explain why it's good. This is critical. AI learns faster from annotated examples than from examples alone. Annotation tells AI what to pay attention to.

4. Varied but Consistent
Different content, but same voice. The reviews should be about different people with different accomplishments, but the tone and structure should feel like they came from the same writer. This teaches AI your signature style across different contexts.

The Annotation Pattern

For each example, add notes explaining why it works.

Format:

"Example [#]: [Title or description]

[Paste the example here]

Why this example works:
- [Reason 1]: [Explanation]
- [Reason 2]: [Explanation]
- [Reason 3]: [Explanation]

Key characteristics:
- Length: [X words/pages]
- Tone: [Description]
- Structure: [How it's organized]"

Real example for performance reviews:

Example 1: Strong Performer, Technical Depth

[Paste a real performance review here that you're proud of]

Why this example works:
- Specific accomplishments (shipped X feature, reduced query time from 2s to 500ms, mentored 3 junior engineers)
- Concrete metrics where possible (30% improvement, 4 people, $200K cost savings)
- Clear growth area (communication with non-technical teams; here's specific area to work on)
- Development action that flows from growth area (join quarterly product strategy meetings; leads to better cross-team communication)
- Length: ~400 words. Tight but comprehensive. Long enough to be substantive, short enough to read in 5 minutes.
- Tone: Direct and appreciative. Not gushing. Not corporate. Sounds like a real human.
- Structure: Accomplishments → Growth → Development Action → Overall Assessment. Clear flow.

Now when you ask AI to draft a similar review, you've trained it not just on content but on the exact structure and voice you want. You've shown it what you mean by "balanced" and "specific."

Using Few-Shot for Performance Reviews

This is where few-shot shines. Performance reviews are where voice and tone matter most.

Step 1: Collect 3-5 Reviews You Like

Pull reviews you've written that you're proud of. Different scenarios:
- High performer (achieved goals, grew, led)
- Solid middle performer (met expectations, steady contributor)
- Someone with clear growth areas (capable but needs development)
- Someone transitioning roles (new to role, ramping up, showing promise)
- Someone new to the company (first year, learning, adjusting)

Step 2: Annotate Each

For each review, add notes:
- Why is it good?
- What specific accomplishments are mentioned?
- How is the growth area handled (honest? constructive? specific?)?
- What's the structure?
- What's the tone?
- What's the length?

Step 3: Prompt AI

"Here are examples of performance reviews from [Company]. Study these. Note the structure, tone, specificity level, balance of accomplishment and growth areas.

[Paste 3-4 annotated reviews]

Now draft a performance review for [Person] who:
- [Role and level]
- [Key accomplishments this year]
- [Growth areas]
- [Strengths]

Use the same structure and tone as the examples. Make it specific, balanced, direct, and constructive."

AI produces something that matches the style of your examples. Not perfect, but much better than generic output.

Using Few-Shot for Recruiting Emails

Recruiting is where few-shot really shines. Recruiting emails are where personalization and voice matter.

You want recruiting emails that:
- Feel personal, not templated
- Are warm but professional
- Mention something specific about the candidate
- Explain why they'd care about YOUR role
- Don't feel salesy
- Sound like you, not like a recruiter bot

Instead of describing this, show examples.

Step 1: Collect Emails You're Proud Of

Pull recruiting emails you've sent that got strong response rates. Different types:
- Outreach to a passive candidate (someone not looking)
- Outreach to someone from a competitor (poaching)
- Outreach to someone from a different industry (career changer)
- Follow-up to someone who didn't respond first time
- Response to someone who reached out to you

Step 2: Annotate

For each email, note:
- Why did this get a response?
- What made it feel personal?
- What specific detail about the candidate is mentioned?
- Why would they care about your company?
- What's the length?
- What's the tone?

Step 3: Prompt AI

"Here are recruiting emails from our team. They're based on actual outreach that got strong response rates. Study them. Note the tone, the personalization, how we explain why someone would care about our role.

[Paste 3-4 annotated emails]

Now draft a recruiting email for [Candidate Name]. Here's what I know about them: [background, achievements, interests]. Here's why I think they'd be interested in [role]: [reasons]. Draft an email using the same tone and approach as the examples."

AI learns the pattern and produces something that feels like your outreach, not a template.

Using Few-Shot for Policy Clarification

You want policy explanations that are clear and simple, not legal.

Collect Examples:

Paste 2-3 policy clarifications you've written:

Bad policy language → Your clear explanation

Example:
- Bad: "Employees may be required to notify their manager of any medical conditions necessitating reasonable accommodation pursuant to the Americans with Disabilities Act."
- Your clear version: "If you have a medical condition that affects your work, let your manager know. We'll work together to figure out what would help you succeed."

Annotate:

"This explanation works because:
- Uses plain language (no 'necessitating,' no 'pursuant')
- Explains the real action (tell your manager)
- Emphasizes collaboration (we'll work together)
- Ends with positive outcome (succeed)
- No legal jargon or passive voice"

Prompt AI:

"Here are examples of policy language + my clear explanations:

[Paste examples with explanations]

Here's another confusing policy: [policy]. Clarify it in the same style as these examples: clear, plain language, collaborative tone, no jargon."

AI learns your voice and produces clarity instead of corporate-speak.

Using Few-Shot for Tone and Voice

This is where few-shot is most powerful.

You can't describe your company's tone perfectly. But you can show it.

Collect Examples:

Gather your best examples of company voice:
- An internal email about a difficult topic (layoffs, strategy change, policy change)
- An email about good news (funding, new product, hiring)
- A policy clarification
- A difficult individual conversation (feedback, performance, issue)

These show how your company talks in different situations.

Annotate:

"This email works for our culture because:
- We're direct but not harsh
- We explain why (context matters)
- We acknowledge difficulty
- We don't sugarcoat
- We assume intelligence and good faith"

Prompt AI:

"Here are examples of how we communicate at [Company]:

[Paste 4-5 annotated examples in different contexts]

I need to communicate: [situation]. Draft something in the same tone as these examples."

AI learns: are you corporate or casual? Honest or diplomatic? Warm or direct? It learns by example, not by description. Examples are infinitely more powerful than descriptions.

Building Your Few-Shot Library

Over time, collect examples:

For recruiting:
- 5 good recruiting emails (that got responses)
- 3 good job descriptions
- 2 good offer letters

For performance management:
- 5 good performance reviews (different scenarios)
- 2 good difficult conversations (performance issue, behavioral issue)
- 3 good goal-setting frameworks

For communications:
- 3 policy clarifications (different types)
- 2 difficult internal communications (bad news, strategy change)
- 3 good one-pager explanations

For general HR:
- 3 examples of your tone in different contexts
- 2 examples of clear vs. unclear writing on same topic

Save these. Organize them. When you need to create something new, pull from your library.

"Here are examples of our recruiting emails. Draft a similar email for [candidate]."

You're not starting from scratch. You're teaching AI your style through examples.

Annotation Quality Matters

Good annotations make examples work better.

Bad annotation:
"This email is good."

Good annotation:
"This email works because:
- It opens with a specific detail about the candidate (they worked at X, which is relevant because we're building Y)
- It explains what we're building in one sentence (clear value proposition)
- It mentions the role without overselling (confident but not salesy)
- It closes with a clear next step (let's grab coffee, not 'let me know if interested')
- Length: 100 words. Long enough to be personal, short enough to read in 30 seconds.
- Tone: peer-to-peer, not recruiter-to-candidate. Conversational."

The second annotation teaches AI not just what's good, but why. It teaches AI what to pay attention to.

Testing Your Examples

Before you deploy a few-shot set, test it.

Ask AI to draft something using your examples. Does the output match the style of the examples? Does it feel like yours?

If yes, great. Use it.

If no, either:
1. Your examples aren't clear enough (improve annotations)
2. Your examples aren't truly representative (pick better ones)
3. Your examples are inconsistent (pick examples that are more similar in voice)

Refine and test again.

Updating Your Library

Good few-shot examples evolve.

If you write something better, replace the old example. If you discover an example doesn't work well as a template, remove it. Your library should show your best work, not your historical work.

Every few months, review your library: "Do these examples still represent the quality and style I want?"

As you improve, your examples should improve too.

Try This Now: Two Hands-On Exercises

Exercise 1: Collect and Annotate Examples

Pick one task you do regularly (performance reviews, recruiting emails, policy clarifications, etc.).

Over the next week, save 3-5 pieces of work you do in that category that you're proud of.

Annotate each: Why is it good? What makes it work? What's the structure, tone, length?

Exercise 2: Test Few-Shot Prompting

Using your 3-5 examples, prompt AI: "Here are examples of [task] from my company. [Paste examples with annotations.] Now draft a similar [task] for [new situation]."

Compare the output to generic AI output. Is it better? More in your style? Does it match your examples?

Practical Application - "What to Do Monday Morning"

• 
  Identify your top 5-10 tasks: What do you do regularly? What would benefit most from style consistency?
• 
  For each task, collect examples: 3-5 examples of your best work.
• 
  Annotate your examples: Why is each one good? What makes it work? What are the key characteristics?
• 
  Organize your library: Folder structure by task. Easy to find.
• 
  Document your library: What are your examples for each task? Where are they stored? How do people access them?
• 
  Use examples in prompts: When asking AI to create something, provide examples first. "Here are examples of X from our company. Draft a similar X for [new situation]."
• 
  Refine your library: As you write better work, replace old examples. Your library should always show your best work.
• 
  Share with your team: Let them use your examples too. This builds consistency and collective quality.

Key Takeaways

• Examples teach better than descriptions: Show AI your style; don't describe it. Seeing is understanding.
  - Annotate your examples: Explain why each one works. Annotation teaches AI what to pay attention to.
  - Use different scenarios: Show examples across different situations. Variety teaches the underlying pattern.
  - Build a library: Save good examples for reuse. This is your style guide in example form.
  - Test your examples: Does output match the style of your examples? If not, improve annotations or choose better examples.
  - Tone and voice are learned through example: This is where few-shot is most powerful. Show, don't tell.
  - Update your library: Remove examples that don't work, add new good ones. Your library should always show your best work.

FAQ

Q: How many examples do I need?
A: 3-5 is typical. Enough to establish a pattern, not so many it's overwhelming. More than 10 usually doesn't improve output quality much. Quality of examples matters more than quantity.

Q: Can I mix examples from different people?
A: Yes, if they're in your company style. But better if they're your own writing. Your style is clearer than "company style amalgamated from many writers."

Q: Should I update my examples?
A: Yes. If you write something better, replace the old example. Your library should show your best work, not your historical work.

Q: What if I don't have good examples yet?
A: Write some. Create 3-5 pieces of work in the task you care about. Make them your best work. Then use them as examples. This also helps you clarify what you actually want.

Q: Can I use few-shot prompting for everything?
A: No. Strategic/conceptual work doesn't need examples. But anything where style, voice, or tone matters, recruiting, communications, reviews, benefits from few-shot. Tasks where accuracy and facts matter less than style benefit most.

Q: What if my examples are inconsistent with each other?
A: That's a problem. Pick examples that are more similar in voice and style. Few-shot works best when examples are consistent in tone, length, and approach.

Q: How do I know if I'm annotating well?
A: If the AI output matches your examples, you're annotating well. If it doesn't, improve annotations. Also: do your annotations help YOU understand why the example is good? If not, improve them.

What's Next

Lesson 8.3 is about structured output: getting AI to produce information in formats you can actually use, tables, checklists, scorecards, matrices, instead of prose. You've learned to guide style with few-shot. Now learn to guide format with structured output prompts.