1. Core Principles (for every prompt)
Start from the user
Open with an observed pain, goal, or context clue.
Prove you understand their world before you pitch solutions.
Make your product the bridge
Immediately tie the pain to a specific capability in your product.
Avoid vague “AI copilot” or “magic AI” language.
Stay warm and human
Confident, plain-English tone.
Contractions are welcome.
No jargon unless the prospect already uses it.
Keep it concise
Aim for 2–4 short paragraphs or bullets.
Trim filler words.
Every sentence should add new information.
Respect channel norms
LinkedIn: no signatures.
Email: use simple, professional signoffs.
Follow-up messages: no hard CTAs if it is meant to be gentle.
Avoid spam triggers
Skip words like “free”, “guarantee”, “urgent”.
No exclamation marks or ALL CAPS.
Personalize responsibly
Only use data points you can prove (funding, role change, recent usage, public info).
Never fabricate specifics.
Guardrails
Do not promise outcomes, integrations, or timelines you cannot guarantee.
No medical or financial advice.
Keep data privacy hints generic:
Use “based on your recent usage trends”, not “based on your Salesforce export on 10/14”.
2. AI Assistant (Katie-style) – Channel Specific Guidance
(“Katie” = your outbound/assistive AI agent – rename as needed.)
A. LinkedIn – First Touch
Goal: Introduce value without pressure.
Lead with a real, likely pain (for example: reps juggling inbox + LinkedIn manually, no time to personalize).
Explain concretely how your product helps:
Surfaces the right lead context.
Drafts replies and outreach steps.
Logs activity automatically.
Tone: helpful, confident, optional.
Do not close with a question.
End with something like: “Happy to chat if you’re exploring this.”
B. LinkedIn – Follow Up
Goal: Re-engage politely and stay relevant.
Acknowledge that you’re following up, without guilt or pressure.
Briefly remind them why you reached out:
Tie to their role, team, or process.
Keep it short and light.
No hard CTA, no “just bumping this”, no pressure language.
C. Email – Subject Lines
Tie directly to the body (for example: “Coordinating LinkedIn + inbox”, “Scaling outbound without more reps”).
Keep it short and natural.
Avoid spam triggers and clickbait.
D. Email – Body
Open with strong personalization:
Company, role, recent change, or a clear situation they’re in.
In 1–2 sentences, connect:
Their pain → how your product solves it.
Keep copy concise, warm, and specific to a workflow (not an abstract “AI” pitch).
No titles or names inside the body if the signature already includes them.
Always close with:
Best regards,
3. Common Do and Don’t
Do:
Cite specific workflows and outcomes:
Sequencing LinkedIn steps.
Drafting and logging replies automatically.
Detecting when audiences or campaigns are exhausted.
Use real, provable triggers:
Funding, hiring, tech stack, public metrics, product usage.
Don’t:
Make tracking claims:
No “I saw you opened/clicked…”.
Overpromise automation:
No “zero manual work ever”.
Mention internal tools, experiments, or implementation details.
4. Call Coaching & Knowledge Prompts (Compass-style)
Use cases
Real-time coaching cards.
Call recaps and summaries.
Objection handling suggestions.
Knowledge search and answer retrieval.
Inputs hierarchy
When giving context to the model, prioritize:
Call transcript snippets (customer’s exact words).
CRM context (deal size, stage, role).
Knowledge base excerpts.
Tell the model what to favor, for example:
“Favor the customer’s exact words over CRM notes when summarizing objections.”
Command style
Use verb-first instructions:
“Summarize objections.”
“Rank follow-ups by risk.”
“Draft 3 ways to respond to this objection.”
Specify desired output shape:
Bullet count.
Tables.
Short paragraph vs playbook style.
Formatting
Ask for bullets or tables explicitly:
“Return 3 bullets.”
“Create a 2-column table: objection vs suggested response.”
Safety
Explicitly forbid medical or legal guidance:
“If the topic is medical or legal, do not provide advice. Ask the rep to consult a professional.”
Evaluation
Require confidence statements:
“If confidence is below 0.7, ask for clarification or show 2–3 options instead of one definitive answer.”
Internal tagging
Include tags for analytics and debugging:
[coach_card|objection][recap|next_steps]
5. Sequences and Outreach Variants
Goal: Keep multi-step prompts consistent, comparable, and testable.
Template layering
Define a base persona/system prompt:
Voice, guardrails, positioning.
Add per-step modifiers:
Day, channel, tone, and objective.
This avoids copy drift between steps.
State awareness
Remind the model what has already been sent:
“Previous step: value introduction.”
“This step: share social proof and one clear CTA.”
Testing hooks
Add experiment tags in the prompt:
[exp=li-soft-open-001]Log variant="social-proof-short"
These can be used downstream for analytics.
Fallbacks
Define behavior when personalization data is missing:
“If there is no recent activity, use a role-based hook instead of company news.”
“If title is unknown, default to ‘sales leader’.”
Compliance
Include compliance reminders:
Unsubscribe or opt-out language where needed.
Respect regional norms (for example, no overly casual tone in conservative markets).
Keep one clear CTA per email.
Don’t mix “book a demo” and “reply with feedback” in the same step.
6. Internal Appendix (For Your Team Only)
(Adapt names/channels to your org.)
Owners
Outbound & AI assistant prompts: Demand Gen / Growth team
Contact: e.g.
#demand-gen
Call coaching / knowledge prompts: Solutions / RevOps / SE
Contact: e.g.
#ai-coaching
Sequences framework: Lifecycle / Marketing Ops
Contact: e.g.
#lifecycle-lab
QA and Sign-Off Flow
Draft in your Notion (or similar) template:
Prompt Playbooks / <Product>Run any lint/check script (if exists):
/prompt-lint run <doc-url>Peer review in Slack/Teams thread.
PM or owner approval.
Publish:
External knowledge base for customer-facing versions.
Internal workspace (Notion, Confluence, Guru) for internal use.
Experiment Tracking
Log in your experiment tracker (Airtable, Notion, sheet), including:
Variant name.
Hypothesis.
Start/end dates.
Metrics: reply rate, positive sentiment, compliance incidents.
Use UTM-like tags inside prompts:
[exp=li-soft-open-001],[exp=email-social-proof-002].
Escalations
Urgent fixes (compliance or factual inaccuracies):
Page your prompt-hotfix / on-call channel.
System bugs:
File in your “Prompt Quality” or AI-related project in Jira/Linear with priority P1–P3.