enrichment-design

Data Points Builder

Bridge the gap between research hypotheses and table enrichment. Define WHAT to research about each company before running enrichment.

When to Use

• After market-research has produced a hypothesis set
• Before list-enrichment — this skill designs the columns, that skill runs them
• When the user says "what should we research about these companies?"

Two Modes

Mode 1: Segmentation

Goal: Design columns that score or confirm hypothesis fit per company.

Input: Hypothesis set (from market-research or context file)

Process:

1. Read the hypothesis set
2. For each hypothesis, propose 1-2 columns that would confirm or deny fit
3. Discuss with user — refine, add, remove
4. Output final column_configs

Example: If hypothesis is "Database blind spot — 80-90% of targets invisible to standard tools":

• Column: "Data Infrastructure Maturity" (select: ["No CRM", "Basic CRM", "Full stack"])
• Column: "Digital Footprint Score" (grade: 1-5)

Mode 2: Personalization

Goal: Design columns that capture company-specific hooks for email personalization.

Input: Target list + what the user wants to personalize on

Process:

1. Ask what hooks matter for this campaign (leadership quotes, recent launches, hiring signals, tech stack, etc.)
2. Propose 2-4 columns with prompts
3. Discuss with user — refine
4. Output final column_configs

Example: For personalization hooks:

• Column: "Recent Product Launch" (text: describe any product launched in last 6 months)
• Column: "Leadership Public Statement" (text: find a public quote from CEO/CTO about [topic])

Interactive Column Design

Do NOT just generate columns silently. Walk through this with the user:

Step 1: Present the framework

Show the user the two modes and ask which applies (or both).

Step 2: Propose initial columns

Based on hypotheses or user input, propose 3-5 columns. For each, show:

Column: [name]
Type: [output_format]
Agent: [research_pro | llm]
Prompt: [the actual prompt text]
Why: [what this tells us for segmentation/personalization]

Step 3: Refine together

Ask:

• "Any columns to add?"
• "Any to remove or merge?"
• "Should any prompts be more specific?"

Step 4: Confirm column budget

Guidance:

• 3-5 columns is the sweet spot
• 6-7 is acceptable if each serves a clear purpose
• 8+ adds noise — push back and suggest merging

Step 5: Output column_configs

Generate the final column configs as a JSON array ready for list-enrichment:

[
  {
    "kind": "agent",
    "name": "Column Display Name",
    "key": "column_key_snake_case",
    "value": {
      "agent_type": "research_pro",
      "prompt": "Research prompt using {input} for domain...",
      "output_format": "text"
    }
  }
]

Column Design Guidelines

Agent Type Selection

Data point type	Agent type	Why
Factual data from the web (funding, launches, news)	research_pro	Needs web research
Classification from company profile	llm	Profile data is enough
Nuanced judgment (maturity, fit score)	research_reasoning	Needs chain-of-thought
People/org structure	linkedin	LinkedIn-specific

Output Format Selection

Data point type	Format	When
Free-form research	text	Open-ended questions
Score/rating	grade	1-5 scale assessments
Category	select	Mutually exclusive buckets
Multiple tags	multiselect	Non-exclusive tags
Structured data	json	Multiple related fields
Yes/no with evidence	json	{"match": bool, "evidence": str}

Prompt Writing Tips

• Always include {input} for the company domain
• Be specific about output format in the prompt itself
• Include fallback: "If not found, return N/A" or "If unclear, return 'Unknown'"
• For select/multiselect: list the labels in the prompt too
• For hypothesis scoring: reference the specific hypothesis in the prompt
• Keep prompts under 200 words

Reference Library

See references/data-point-library.md for ~20 pre-built column configs organized by use case.

Output Handoff

After column design is complete:

1. Present the final column_configs JSON to the user
2. Tell the user: "These configs are ready for list-enrichment. Run that skill with your table ID and these columns."
3. If the user wants to run immediately, hand off to list-enrichment workflow