clawd - OpenClaw Agent Workflow

This repository defines an OpenClaw-friendly agent workflow for prompt-first lead intelligence, JSON auditability, and HubSpot handoff design.

Tags: openclaw ai-agents multi-agent prompt-engineering lead-intelligence hubspot json-workflows

The workflow is built for:

1. finding potential leads from LinkedIn profiles
2. finding pain signals from LinkedIn posts and comments
3. qualifying strong opportunities
4. syncing qualified leads into HubSpot
5. storing research data locally as JSON

The repository is intentionally prompt-first. Each agent has a clear contract, a fixed input, a fixed output, and a narrow responsibility.

Pipeline Summary

There are two collection paths and one shared CRM destination.

Lead pipeline

1. agents/linkedin_scout/PROMPT.md Scrape founder/operator profiles that show pain around PRDs, planning, shipping, or AI coding workflows.
2. agents/hubspot_processor/PROMPT.md Normalize, dedupe, score, tier, and upsert strong profile leads into HubSpot.

Content pipeline

1. agents/content_scout/PROMPT.md Scrape public LinkedIn posts and comments that contain pain signals.
2. agents/insight_storer/PROMPT.md Store every content insight locally, dedupe it, and promote only strong authors into HubSpot.

Orchestration

agents/orchestrator/PROMPT.md defines the run order, handoff rules, and cron behavior.

Design Principles

• Scouts only collect evidence. They do not score, message, or write to HubSpot.
• Processors decide. They normalize, dedupe, score, tier, and sync.
• Every run should be idempotent. Running the same cron twice should not create duplicate HubSpot contacts or duplicate local insights.
• Local JSON is the system of record for raw research. HubSpot is the system of record for qualified leads.
• Prompts should be readable by a human reviewer without needing hidden context.

Recommended Cron Setup

Use two cron jobs instead of one large mixed run.

Cron 1: lead discovery

Runs the orchestrator in lead mode:

1. run linkedin_scout
2. verify data/leads/raw_candidates.json exists and is valid
3. run hubspot_processor
4. verify data/leads/qualified_leads.json exists and summarize results

Cron 2: content discovery

Runs the orchestrator in content mode:

1. run content_scout
2. verify data/content/content_insights.json exists and is valid
3. run insight_storer
4. verify data/insights.json exists and summarize results

This separation keeps failures isolated and makes reporting easier.

Data Contracts

Canonical output files:

• data/leads/raw_candidates.json
• data/leads/qualified_leads.json
• data/content/content_insights.json
• data/insights.json

Each file should contain:

• runId
• capturedAtISO or processedAtISO
• source
• a top-level array of records

Each record should contain enough evidence to justify why it exists.

What A Team Lead Should Expect

This repo is not a generic scraping folder. It is a controlled lead-intelligence workflow with:

• narrow agent ownership
• explicit handoffs
• JSON-first auditability
• HubSpot sync guardrails
• no automated outreach sending

Immediate Next Implementation Layer

If you want this to become executable rather than prompt-only, the next layer should be:

1. scripts/validate_json.*
2. scripts/hubspot_upsert.*
3. scripts/run_pipeline.*
4. optional schemas/*.json
5. logs/ or data/runs/ for archived run manifests

The prompts in this repository are now structured so those scripts can be added without changing the business flow.