GTM Intelligence Agent

An agent that researches a company and a job posting, then generates a structured briefing with talking points and a draft outreach opener — built directly on the Claude API (Python).

Project structure

gtm-agent/
├── gtm_agent.py          # the main tool — run this
├── requirements.txt      # dependencies
├── .gitignore            # keeps API keys and personal briefs out of git
├── archive/              # earlier day-by-day builds, kept for reference
│   ├── day1_basic_call.py
│   ├── day2_web_search.py
│   └── day3_chained_research.py
└── README.md

Setup

git clone <your-repo-url>
cd gtm-agent
pip3 install -r requirements.txt
export ANTHROPIC_API_KEY="sk-ant-your-key-here"

Usage

python3 gtm_agent.py \
  --company "Acme Corp" \
  --role "https://example.com/jobs/some-role" \
  --background "Salesforce Admin, Salesforce Consultant, GTM Systems Engineer, Business Analyst" \
  --years 5

Flags:

• --company (required) — company name
• --role (required) — role title OR a job posting URL
• --background (optional) — your angle, used to tailor talking points
• --years (optional) — avoids placeholder text in the outreach opener
• --output-dir (optional) — where to save the briefing (default: briefs/)
• --no-save (optional) — print only, skip saving to file

Each run saves a timestamped .md briefing, e.g. briefs/2026-07-06_acme-corp_brief.md.

How it works

1. research_company() — uses web_search to pull current intel on the company: business model, core products, recent strategic moves, competitive pressure
2. research_role() — behavior depends on what you pass as --role:
   • Job posting URL → uses web_fetch to read the full page directly, getting the actual job description rather than search results about it. Falls back to web_search if the page is behind a login wall.
   • Role title text → uses web_search to research what the role typically involves and requires
3. synthesize() — a non-search call that reasons over steps 1 and 2 and produces one structured brief: Company Snapshot, Role Fit Signals, Talking Points, Draft Outreach Opener

This is "programmatic chaining" — the workflow shape and what data flows between calls is deliberately architected, not left to one freeform prompt. Each call uses the right tool for the job (web_fetch vs. web_search), which is a meaningful architectural choice, not just a different API call.

Known limitations

• Login-walled job postings (Workday, Greenhouse, Lever behind SSO) — web_fetch will fall back to web_search automatically, but the result will be less specific than a direct page read. In practice, postings on company career pages or LinkedIn tend to work well.

Build history

Stage	What changed
Day 1	Basic Claude API call
Day 2	Added web_search tool use
Day 3	Chained company research → role research → synthesis
Day 4	Wrapped as a real CLI tool (argparse, file output, --flags)
Day 5	Swapped web_search for web_fetch on job posting URLs for full-page content

The archive/ folder holds the day-by-day scripts for reference. gtm_agent.py is the current production version.