audience-builder（观众建设者）

Audience Builder

Purpose

The primary end-to-end skill. Takes a natural language audience description and produces a validated, sized, and confirmed segment. Composes schema-discovery, filterql-builder, segment-manager, and confirmation-gate.

Environment

• LYTICS_API_TOKEN -- API authentication token
• LYTICS_API_URL -- Base URL (default: https://api.lytics.io)

Inputs

• Natural language description of the desired audience
• Optional: segment name, table (default: user), kind (default: segment), tags

End-to-End Flow

Step 0: Check if Advisor Would Be Better

If the user's request is goal-oriented rather than specific (e.g., "I want to drive more conversions" vs "US users who visited 5+ times"), suggest starting with the audience-advisor skill first:

"It sounds like you have a business goal in mind. Would you like me to analyze your data first to recommend the best targeting strategy? I can look at ML models, field distributions, and existing segments to suggest evidence-based filters."

If the user has a specific audience description, proceed directly to Step 1.

Step 1: Parse Intent

Extract the key concepts from the natural language description:

• Target attributes (demographics, behaviors, properties)
• Inclusion criteria (what they have/did)
• Exclusion criteria (what they lack/didn't do)
• Temporal constraints (time-based conditions)

Example: "US customers who haven't purchased socks in over a year"

• Attribute: country = US
• Exclusion: no sock purchases
• Temporal: over 1 year ago

Step 2: Schema Discovery

Use schema-discovery skill to find matching fields:

# Get all fields to find candidates
curl -s "${LYTICS_API_URL:-https://api.lytics.io}/v2/schema/user/field" \
  -H "Authorization: ${LYTICS_API_TOKEN}"

Score candidate fields by name match, type compatibility, and value distribution. For the example:

• "US customers" -> look for: country, geo_country, country_code (string type)
• "purchased socks" -> look for: products_purchased (set), purchase_categories (set), last_purchase_date (date)

Fetch value suggestions for top candidates -- the q parameter is required:

curl -s "${LYTICS_API_URL:-https://api.lytics.io}/api/schema/user/fieldsuggest/${FIELD}?q=${QUERY}" \
  -H "Authorization: ${LYTICS_API_TOKEN}"

Example: /api/schema/user/fieldsuggest/country?q=US

Confirm field values match (e.g., country contains "US", products_purchased contains "socks").

Step 3: Build FilterQL

Use filterql-builder skill to construct the filter expression.

Map each concept to a FilterQL condition based on field type (see ../references/field-types.md):

• String equality: country = "US"
• Set non-membership: NOT products_purchased INTERSECTS ("socks")
• Date comparison: last_sock_purchase < "now-1y"

Compose with AND/OR:

FILTER AND (
  country = "US",
  NOT products_purchased INTERSECTS ("socks")
) FROM user ALIAS us_no_socks

Step 4: Validate FilterQL

curl -s -X POST "${LYTICS_API_URL:-https://api.lytics.io}/api/segment/validate" \
  -H "Authorization: ${LYTICS_API_TOKEN}" \
  -H "Content-Type: text/plain" \
  -d 'FILTER AND (country = "US", NOT products_purchased INTERSECTS ("socks")) FROM user ALIAS us_no_socks'

If validation fails:

1. Parse the error message
2. Identify the syntax issue
3. Fix the FilterQL
4. Retry (max 3 attempts)
5. If still failing, show the error and ask the user for guidance

Step 5: Estimate Size

curl -s -X POST "${LYTICS_API_URL:-https://api.lytics.io}/api/segment/size" \
  -H "Authorization: ${LYTICS_API_TOKEN}" \
  -H "Content-Type: text/plain" \
  -d 'FILTER AND (country = "US", NOT products_purchased INTERSECTS ("socks")) FROM user ALIAS us_no_socks'

Step 6: Present Draft (Confirmation Gate)

Show the user a complete summary before creating:

## Proposed Audience: US Non-Sock Buyers

**Name**: US Non-Sock Buyers
**Slug**: us_no_socks
**Table**: user
**Kind**: segment

**Description**: US customers who have not purchased socks

**Field Mapping**:
- "US customers" -> `country = "US"` (field type: string, sample values: US, CA, UK)
- "haven't purchased socks" -> `NOT products_purchased INTERSECTS ("socks")` (field type: []string)

**FilterQL**:
FILTER AND (country = "US", NOT products_purchased INTERSECTS ("socks")) FROM user ALIAS us_no_socks

**Estimated Size**: ~45,000 profiles

Then show the raw API payload:

{
  "name": "US Non-Sock Buyers",
  "slug_name": "us_no_socks",
  "description": "US customers who have not purchased socks",
  "segment_ql": "FILTER AND (country = \"US\", NOT products_purchased INTERSECTS (\"socks\")) FROM user ALIAS us_no_socks",
  "kind": "segment",
  "table": "user",
  "is_public": true,
  "save_hist": true,
  "tags": []
}

Ask for confirmation: "Proceed with creating this segment? (yes/no)"

Step 7: Execute on Approval

curl -s -X POST "${LYTICS_API_URL:-https://api.lytics.io}/v2/segment" \
  -H "Authorization: ${LYTICS_API_TOKEN}" \
  -H "Content-Type: application/json" \
  -d '{ ... approved payload ... }'

Report the result:

Successfully created segment "US Non-Sock Buyers" (id: abc123)
View at: ${LYTICS_API_URL}/segment/us_no_socks

After successful creation, offer an audience snapshot: "Segment created successfully. Would you like to see an audience snapshot showing the demographic breakdown and field distributions?" If yes, invoke the audience-snapshot skill with the new segment ID.

Handling Ambiguity

When field mapping is ambiguous (multiple equally strong candidates):

1. Present the top candidates with sample values
2. Let the agent pick the best match based on context -- do NOT prompt the user unless truly ambiguous
3. Only ask the user when there's a genuine tie or the field purpose is unclear

Updating Existing Segments

If the user wants to update an existing segment:

1. Fetch the current segment: GET /v2/segment/${ID}?inline=true
2. Show current FilterQL and proposed changes
3. Use PUT /v2/segment/${ID} instead of POST

Error Handling

• No matching fields: Report clearly, suggest the user describe their data differently
• Validation failure after 3 retries: Show the FilterQL and error, ask for help
• Empty segment (size 0): Warn and suggest broadening criteria
• Very large segment (>90% of total): Note this and confirm intent

Dependencies

• Composes: schema-discovery skill, filterql-builder skill, segment-manager skill, ../references/confirmation-gate.md
• References: ../references/filterql-grammar.md, ../references/field-types.md