strategy.md

• AI variable names must be Title Case with spaces:

  Subject Line

  ,

  First Line

  ,

  PS Line

  . They become

  [Title Case]

  brackets in emails.md and H3 headings in prompts.md - the three have to agree exactly.
• AI variable names ≤30 chars (enforced in prompts.md by the prompt-writer contract).
• Never reference data sources personalization tools typically lack (news, funding, events, social media, blogs) - these propagate into prompts.md and break at runtime when the data is not available.

1. Read

   {workspace}/research.md

   - Client research: overview, ICP, value prop, case studies, lead magnets
2. Read

   {workspace}/notes.md

   (if it exists) - Client-specific preferences
3. Read plan.md from parent directory - Get segment definition, angle, ICP, and testing hypothesis

{workspace}

./cold-email/{slug}/

{workspace}/
├── research.md       # Client research - ICP, value prop, case studies
├── notes.md          # Optional client preferences and feedback
└── {plan_folder}/    # One per campaign, e.g. "03-15 - plan 1"
    ├── plan.md           # Plan with ICP, segments, angles, testing hypothesis
    ├── 1 - {segment}/
    │   └── strategy.md   # THIS file - segment strategy
    ├── 2 - {segment}/
    │   └── strategy.md
    └── ...

1. Segment folder path - e.g.,

   {workspace}/03-15 - plan 1/1 - early-stage/

2. Plan.md - Read from parent directory of segment folder to get:
   • Broad ICP definition
   • Segment-specific criteria
   • Messaging angle for this segment
   • Testing hypothesis (if any)
3. Client Research Document - Check

   {workspace}/research.md

   or uploads
4. Campaign Goal - What action to drive:
   • Book demos/meetings
   • Drive traffic to landing page
   • Get replies for qualification
   • Promote lead magnet

• ICP: {Broad ICP from plan.md}
• Segment: {Specific segment criteria from plan.md}
• Angle: {Messaging angle from plan.md}
• Goal: reply-based / click-based
• Funnel Type: {type}

• Campaign uses AI personalization (first lines, PS lines, body variables)
• Higher quality/variation is desired for personalization outputs
• Budget allows ($5 per 1,000 contacts for rewriting flow)

• No AI personalization variables exist
• Subject lines only (rewriting too risky for subjects)
• Budget constraints require minimizing AI costs

1. Building block approach - Treat copy and campaigns as modular building blocks (like Legos) that can be tested in many combinations. Volume of combinations is how we discover what works. See

   references/building-blocks.md

   for the full philosophy.
2. Start strong, diversify, close strong - Lead email uses strongest angle. Middle emails try new angles. Final email is a breakup with second-strongest angle.
3. Relevance over personalization - Don't just use {firstName}. Show deep understanding of their world using AI variables sourced from LinkedIn/website.
4. One idea per email - Each email has one goal, one CTA, one main point.
5. Value-first - Every touch provides insight, education, or genuine help.
6. Data constraints - AI variables can ONLY use: company website main page, LinkedIn company page, LinkedIn individual profile. No news, events, funding announcements, or timely triggers.

• 3-5 emails maximum
• Delays: 0 days for first, 2-4 days between follow-ups
• CTAs: Reply-based for meetings, click-based for lead magnets. Every email's CTA asks for the actual offer, including Email 1 (which leads with the strongest version of that ask). No soft diagnostic questions as CTAs. Never send naked booking links - ask a direct question first.
• Length: keep static bodies punchy - at or under ~80 words (follow-ups well under). email-review enforces a hard ceiling.
• Format: AI variables use

  {{Title Case}}

  ; every email signs with

  {sender_signature}

  .

1. Receive segment folder path
2. Read plan.md from parent directory to get ICP, segment definition, and angle
3. Load client research from

   {workspace}/research.md

4. Read references:
   • references/building-blocks.md - Campaign testing philosophy
   • references/email-sequence-guide.md - Sequence planning details
   • references/ai-variables-guide.md - Personalization patterns
   • references/sequence-tests-guide.md - Email sequence A/B testing
5. Check for test opportunities (optional):
   • Read plan.md Testing Hypothesis section
   • Evaluate if sequence tests would SUPPORT (not conflict with) macro test
   • Only propose tests that refine HOW, not change WHAT (ICP/angle/offer)
   • Skip tests if first campaign for client, low TAM, or no clear hypothesis
6. Generate strategy document with email sequence (Markdown) and AI variables (JSON)
7. Save to

   {segment_folder}/strategy.md

• Value props from client research are woven through sequence
• Segment-specific pain points inform email hooks
• Angle from plan.md is reflected in the messaging
• Funnel type matches campaign goal
• Lead magnets integrated at appropriate point
• AI variables feasible with available data sources
• Each email has distinct purpose in sequence progression

• mode: "edit"

  indicates edit mode

• trigger: "analytics"

  or

  trigger: "feedback"

  indicates source

• improvement_instruction

  contains specific changes to make

1. Read improvement context: Extract

   improvement_instruction

   object
2. Load existing strategy.md: Parse current strategy from segment folder
3. Identify preservation zones: What should NOT change (see below)
4. Identify change zones: What specifically to modify
5. Apply scoped changes: Only modify specified sections
6. Generate diff: Show before/after for changed sections
7. Get approval: Wait for user to confirm changes
8. Save: Update strategy.md with approved changes
9. Cascade check: Notify if changes affect downstream files (emails.md, prompts.md)
10. Log: Update improvement-log.md in plan folder

• ICP and segment definition (unless pivoting)
• Available data sources and constraints
• Funnel type (reply-based vs click-based)
• Core value proposition from client research

• Email sequence structure if only changing copy angle
• AI variable names if prompts are working

• Email descriptions and goals
• AI variable definitions
• Messaging angle and hooks
• Offer/CTA specifics

• primary_goal: Re-engage with social proof
• description: Share a case study relevant to their industry
• delay_in_days: 3

• primary_goal: Re-engage with curiosity hook
• description: Ask a thought-provoking question about their growth strategy, no case study yet
• delay_in_days: 3

• emails.md: Email 3 will need rewriting (invoke email-copywriting after approval)
• prompts.md: No changes needed (same AI variables)

undefined

analytics_context

• Use

  sequence_stats

  to identify which emails need strategic changes
• Reference

  weak_spots

  for specific improvement areas
• Consider

  sibling_comparison

  to learn from winning campaigns

• Low Email 1 opens = Strengthen opening hook strategy
• Drop-off after Email 2 = Email 3 strategy needs pivot
• Low replies across all = CTA or offer strategy issue

feedback_context

• "Wrong angle" = Pivot messaging strategy
• "Offer not resonating" = Update lead magnet/CTA strategy
• "ICP mismatch" = Update segment overview section
• "Emails too long" = Adjust email descriptions

• Only specified sections were modified
• Preservation zones respected
• Changes align with client research
• Cascade impacts identified
• Diff shown and approved by user

Two A/B levels - don't confuse them. This skill defines email-step A/B (

### Variant A/B

inside one step - subject, hook, CTA). Testing whole sequences against each other is a different thing: a segment-level sequence A/B test, set up in campaign-planning as

percentage

mode (random %-distribution arms, each arm a totally different sequence). If the plan is in

percentage

mode, each arm is its own segment folder and you write its full sequence normally - the "test" is the arm itself, so you usually skip within-step variants in that arm.

• ICP testing (different personas/industries)
• Messaging Angle testing (different hooks/pain points)
• Lead Magnet testing (different offers)
• Whole-sequence testing (

  percentage

  mode - each arm a different full sequence)

• Email execution tests within a SINGLE segment/arm
• Tests defined here, implemented in emails.md by email-copywriting skill
• Variants appear in same emails.md file (Variant A / Variant B)

• Plan has clear macro hypothesis to support (sequence test refines delivery)
• Campaign targets 5,000+ prospects (enough volume to split 50/50)
• Two reasonable approaches exist for an element
• Client expressed preference worth validating
• Historical data suggests optimization opportunity

• Macro test is already complex (don't add noise)
• Limited TAM (<2,000 contacts - not enough volume)
• First campaign for client (establish baseline first)
• No clear hypothesis (testing for testing's sake)
• Tests would conflict with plan's macro test variable

1. Maximum 1 test per strategy (keep it simple)
2. Must reference plan.md's test variable (show how it supports the macro test)
3. Hold constant: ICP, messaging angle, offer, AI variables
4. Vary only: Delivery elements (subject format, PS line, body length, CTA phrasing, timing)
5. Clear hypothesis: State expected outcome and why
6. Measurable: Define primary metric (open rate / reply rate / click rate)

• Variant A (Control): [Default approach - describe]
• Variant B (Test): [Alternative approach - describe]
• Hypothesis: [Why we think B might outperform A]
• Primary Metric: [Open rate / Reply rate / Click rate]
• Affected Emails: [Which emails have variants - e.g., "Email 1 only", "All emails", "Emails 2-3"]

undefined

• Variant A (Control): Include AI-personalized PS line after signature
• Variant B (Test): No PS line - end email at CTA
• Hypothesis: The guarantee message is already strong; PS line may dilute focus or add unnecessary length
• Primary Metric: Reply rate
• Affected Emails: Email 1 and Email 4 (the emails that currently use {{PS Line}})

• Email 1: Write two versions - one with {{PS Line}} placeholder, one without
• Email 4: Write two versions - one with {{PS Line Final}} placeholder, one without
• Emails 2-3: No variants needed (they don't have PS lines)

undefined