> ## Documentation Index > Fetch the complete documentation index at: https://docs.bitscale.ai/llms.txt > Use this file to discover all available pages before exploring further. # AP-2: Prompt Engineering for Outreach > The quality of your AI-generated copy is entirely determined by the quality of your prompts. Learn the principles, patterns, and testing methodology that produce great outreach copy at scale. **AP-2 · AI Personalization · 125 XP · \~22 min** A bad prompt produces generic output, no matter how much context you provide. A great prompt constrains the AI to produce exactly what you want — the right length, the right tone, the right structure, the right level of specificity. Prompt engineering for outreach is a learnable skill. This module covers the patterns that work. *** ## The Anatomy of an Effective Outreach Prompt Every high-performing outreach prompt has six components: ``` 1. ROLE: Who the AI is writing as 2. CONTEXT: What you know about the prospect 3. OBJECTIVE: What the email needs to accomplish 4. CONSTRAINTS: Length, tone, structure, what to avoid 5. OUTPUT FORMAT: Exactly what to return (just the email body, or JSON, etc.) 6. EXAMPLES: Optional, but dramatically improves quality for nuanced copy ``` *** ## Component-by-Component Breakdown ### 1. Role Definition Don't just say "write an email." Tell the AI who is writing it. **Weak:** > "Write a cold email to this prospect." **Strong:** > "You are writing on behalf of a B2B sales rep at Bitscale. Bitscale is a GTM automation platform that helps sales teams build, enrich, and activate prospect data using AI workflows. You are not writing as a marketer — you write like a thoughtful sales rep who has done research and genuinely believes this is relevant." The role definition sets the voice, credibility, and intent of the email. ### 2. Context Variables Include all the enrichment data you've built in prior tracks: ``` Prospect context: - Name: {{first_name}} {{last_name}} - Title: {{job_title}} - Company: {{company_name}} - Industry: {{company_industry}} - Company size: {{company_size_range}} - Funding stage: {{funding_stage}} - Recent signal: {{signal_description}} - Tech stack maturity: {{tech_stack_maturity}} - Messaging angle: {{messaging_angle}} - LinkedIn activity summary: {{linkedin_summary}} ``` More context = better output, up to a point. Don't include every field — include the fields that are most likely to surface specific, relevant observations. ### 3. Objective Be explicit about what success looks like: **Weak:** "Write a cold email." **Strong:** "Write a cold email that gets a reply. The goal is to get them to agree to a 15-minute conversation. The email should feel like it came from a human who did research — not automation." ### 4. Constraints This is where most prompts are too vague. Be extremely specific: ``` Constraints: - Maximum 80 words. Count them. Reject internally if over limit. - No bullet points in the email body - No use of phrases: "I hope this finds you well", "quick question", "just checking in", "thought leader" - First person singular only — "I" not "we" - One CTA only — do not offer multiple options - Framework: Signal-Value-Close (signal in sentence 1, value in sentence 2, close in sentence 3-4) - No exclamation marks ``` ### 5. Output Format Be explicit about what to return: ``` Return ONLY the email body text. No subject line. No greeting salutation line. No sign-off. Do not include labels, headers, or explanations. ``` ### 6. Examples (When to Include) Include 1–2 examples when: * The tone is nuanced (e.g., "direct but not aggressive") * The structure is unconventional * You've tested variants and know what performs best ``` Example of the voice and style I want: --- Saw you just closed your Series B — timing on this is probably good. We help teams like yours skip the 3-month data infrastructure build that usually delays the first SDR hire. Worth 15 minutes to see what it looks like in practice? --- ``` *** ## Prompt Testing Methodology Great prompts are discovered, not written perfectly on the first try. Here's the testing process: **Step 1: Write v1 prompt** Use the six-component structure above. Run it on 10 sample contacts. **Step 2: Quality review** For each output, score on 1–5: * Specificity (does it feel personal?) * Naturalness (does it sound human?) * Compliance (did it follow all constraints?) * CTA clarity (is the ask clear?) **Step 3: Identify failure modes** Where does the prompt consistently fail? Too long? Generic value prop? Wrong tone? **Step 4: Iterate one variable at a time** Change one thing in the prompt. Re-run on the same 10 contacts. Compare scores. **Step 5: Establish your v2 prompt** After 3–4 iterations, you should have a prompt that scores 4+ consistently. *** ## Common Prompt Failures and Fixes | Failure Mode | Symptom | Fix | | ---------------------- | ------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------ | | Generic output | "Companies like yours often struggle with..." | Add stronger role definition + more specific context variables | | Too long | Consistently 120–150 words | Add explicit word count constraint with instruction to "count internally" | | Hollow personalization | References signal but doesn't connect it to anything specific | Add the "so what" instruction: "connect the signal to the specific downstream problem they're likely experiencing" | | Corporate tone | "I wanted to reach out regarding..." | Add example of desired voice + explicit instruction to avoid corporate phrases | | Wrong structure | Doesn't follow SVC or whatever framework you want | Name the framework explicitly and describe each component in the constraints | *** **Quick Check:** What are the six components of an effective outreach prompt? Why do you include role definition? What is the right testing methodology for prompt iteration? *** ## AP-2 Challenge: Build and Test 3 Prompt Variants (+125 XP) Write 3 variants of a personalized outreach prompt (varying one element between each variant). Test all three on the same 10 contacts. **Requirements:** * All 3 prompts documented with all six components * 10 contacts × 3 variant columns = 30 AI-generated emails * Quality scoring column for each email (specificity, naturalness, compliance, CTA — each 1–5) * Average quality score per variant * Winner analysis: which prompt variant performed best, and why? Share your grid + prompt variants + winner analysis. **+125 XP on approval.** *** AP-3 combines company, role, and individual signals into a single layered personalization system.