Fewshot Attempt 01

HarborCRM Task Solving Skill

Overview

This skill captures recurring patterns for solving HarborCRM CRM-integration tasks. Tasks involve querying a local HarborCRM API, applying business rules, and producing structured JSON output matching a provided answer template.

Environment Setup

• The HarborCRM API is available after running env/setup.sh.
• Base URL is supplied by the runner; default for local runs is http://127.0.0.1:8067.
• Only query endpoint paths and entity IDs explicitly named in the task prompt. Do not use global list/index endpoints to discover IDs.

Core Workflow

1. Read the prompt to identify the task type (event reconciliation, trade-show prospecting, batch import cleanup, etc.).
2. Read the answer template (input/payloads/answer_template.json) to understand the exact required schema, keys, enums, and sort orders.
3. Query only the explicitly named endpoints from the prompt.
4. Apply business rules (qualification, exclusion, CRM action mapping, opportunity sizing, normalization).
5. Sort all lists exactly as specified in the template or prompt.
6. Produce JSON only — no explanatory prose outside the JSON.

Task-Type Patterns

1. Post-Event Reconciliation (e.g., neuralops_2026, edgeai_field_2026)

Data to fetch:

• Event details, orders, badges, sponsor packages
• Finance invoices (filtered by event_id)
• CRM accounts, contacts, opportunities, campaign members
• Policies (for exclusion/qualification rules)

Key business rules:

• Sponsor statuses use controlled values: paid_deferred, open_invoice, proposal_only (and sometimes not_sponsor).
• Revenue totals are integer USD. For open_invoice, report the open balance separately.
• Exclude from lead handoff: sponsor attendees, inactive/canceled sponsors, non-business badges, existing disqualified CRM accounts.
• Qualified leads are non-sponsor accounts with business badges that are not already disqualified.
• Opportunity amount comes from the event's configured lead opportunity amount (e.g., 42000, 18000).
• Follow-up due dates are derived from the event date (e.g., leads due 5 business days after event end; sponsor finance due 3 business days after).
• CRM action counts summarize implied account/contact/campaign-member creates and updates.

Sorting conventions:

• sponsor_statuses: by account_name ascending
• qualified_lead_accounts: by account_name ascending
• excluded_records: by company_name ascending, then contact_name ascending
• badge_decisions: by badge_id ascending
• campaign_member_actions: by subject_key ascending
• badge_only_contacts: by company_name ascending

Normalization rules:

• Email: lowercase, trimmed; empty string when missing
• Phone: digits only; empty string when missing

2. Trade-Show Prospecting (e.g., aquafarm_robotics_2026, marinesense_2026)

Data to fetch:

• Trade-show details (/api/tradeshows/{show_id})
• Exhibitors (/api/tradeshows/{show_id}/exhibitors)
• Meeting interest (/api/tradeshows/{show_id}/meeting_interest)
• CRM accounts, policies

Qualification rules:

• Qualified leads are exhibitors that make or OEM-build robotics or underwater-camera platforms covered by the prospecting policy.
• Exclude distributors, service-only providers, sensor-only vendors, research-only entities.
• Existing CRM accounts should be marked for update_existing; new qualified exhibitors for create_account.
• Excluded exhibitors remain visible with a controlled exclusion_reason and crm_action: no_import.

Ranking rules (prospecting):

1. Demo request first (requested_demo: true before false)
2. Meeting interest score descending
3. Broader platform coverage (more platforms = higher rank)
4. Company name ascending (tie-breaker)

Opportunity sizing by priority tier:

• A: USD 120000 — demo-requested qualified leads with score ≥ 90
• B: USD 90000 — demo-requested qualified leads with score ≥ 80
• C: USD 50000 — all other qualified leads

Platform enums: AUV, ROV, Underwater Camera

• Platforms list must be sorted in the order: AUV, ROV, Underwater Camera

Summary fields:

• qualified_lead_count, excluded_count
• existing_crm_overlap_count and existing_crm_overlap_account_ids (sorted ascending)
• total_estimated_opportunity_usd (sum of lead opportunities)
• platform_coverage_counts: counts of leads covering each platform

Excluded exhibitor schema:

• Required: company_id, company_name, relationship_type, exclusion_reason, crm_action
• relationship_type enums: distributor, service_provider, sensor_vendor, research
• exclusion_reason enums: distributor_only, service_only, sensor_only, research_only
• Sort excluded exhibitors by company_name ascending

3. Batch Import Cleanup (e.g., fall_webinar_import)

Data to fetch:

• Import batch raw contacts (/api/import_batches/{batch_id}/raw_contacts)
• Suppression list (/api/import_batches/{batch_id}/suppression)
• CRM accounts and contacts (for deduplication and existing record matching)
• Policies

Business rules:

• Remove duplicates by email (keep the latest/most complete record as winner).
• Remove suppressed emails/domains.
• Remove unusable rows (missing contact name, missing email, etc.).
• Match remaining rows to existing CRM accounts by company name/email and to existing contacts by email.
• CRM action per row: create_account, update_existing, no_import, suppress.
• If a row matches an existing account but not an existing contact, action is update_existing for account and create_contact for contact.

Output sections:

• clean_contacts: import-ready rows with clean_contact_id, source_row_id, company/contact info, crm_action, existing_account_id, existing_contact_id
• duplicate_summary: count and details of duplicates removed (winner/removed mapping)
• removal_summary: count and reasons for all removed rows (duplicate, suppressed, missing_contact, etc.)
• import_action_totals: aggregate counts per crm_action
• campaign_member_import_count: number of clean contacts that will become campaign members

Universal Conventions

JSON Output Rules

• Return JSON only — no markdown, no prose, no comments.
• Match the answer template schema exactly; include all required keys.
• Use null (not missing keys) for nullable fields unless the template says otherwise.
• Currency amounts are always integer USD.
• Dates are ISO 8601 (YYYY-MM-DD) unless the template specifies otherwise.

Sorting Rules

• Always sort lists as specified in the template or prompt.
• Common sort keys: account_name ascending, company_name ascending, badge_id ascending, subject_key ascending, rank ascending (1-based contiguous).
• For tie-breakers, fall back to secondary keys (e.g., contact_name, company_id).

CRM Actions

Common crm_action values across tasks:

• create_account — new account + contact
• update_existing — update matched account
• create_contact — new contact on existing account
• no_import — excluded or suppressed
• no_action — already correct state
• add_campaign_member / create / update — campaign member actions

Exclusion Reasons

Common exclusion reasons:

• sponsor_attendee
• non_business_badge
• existing_disqualified
• missing_contact
• distributor_only
• service_only
• sensor_only / sensor_vendor_only
• research_only

API Query Discipline

• Only call endpoints explicitly listed in the prompt.
• Do not call global list endpoints (/api/events, /api/tradeshows, etc.) to discover IDs.
• Use query parameters exactly as specified (e.g., ?event_id=neuralops_2026).

Common Pitfalls

1. Missing sort order — Many lists have strict ordering requirements. Check the template.
2. Wrong opportunity amounts — Use the event-specific lead opportunity amount, not a hardcoded default.
3. Platform list order — In trade-show tasks, platforms must be ordered AUV, ROV, Underwater Camera.
4. Duplicate handling — In batch imports, deduplicate by email and report the winner/removed mapping precisely.
5. Normalization — Emails must be lowercase and trimmed; phones digits-only.
6. Null vs missing key — Use null for fields like crm_account_id when there is no match; do not omit the key.
7. Contiguous ranks — Ranks must be 1-based and contiguous (no gaps after filtering).
8. Open invoice balance — For sponsors with open_invoice, calculate open_balance = package_amount - paid_amount.